Zuckerberg rushing into every new fad with billions of dollars has somehow tricked people into thinking that's what big tech is about and all of them should be shovelling money into this.
But actually every other company has been much more strategic, Microsoft is bullish because they partnered up with OpenAI and it pumps their share price to be bullish, Google is the natural home of a lot of this research.
But actually, Amazon, Apple etc aren't natural homes for this, they don't need to burn money to chase it.
So there we have it, the companies that have a good strategy for this are investing heavily, the others will pick up merges and key technological partners as the market matures, and presumably Zuck will go off and burn $XB on the next fad once AI has cooled down.
I generally agree with you, although Amazon is really paranoid about being behind here.
On the last earnings call the CEO gave a long rambling defensive response to an analyst question on why they’re behind. Reports from the inside also say that leaders are in full blown panic mode, pressing teams to come up with AI offerings even though Amazon really doesn’t have any recognized AI leaders in leadership roles and the best talent in tech is increasingly leaving or steering clear of Amazon.
I agree they should just focus on what they’re good at, which is logistics and fundamental “boring”
compute infrastructure things. However leadership there though is just all over the map trying to convince folks their not behind vs just focusing on strengths.
Doesn't Amazon have a huge lead just because of AWS? Every other player is scrambling for hardware/electricity while Amazon has been building out data centers for the last 20 years.
> Doesn't Amazon have a huge lead just because of AWS?
They have huge exposure because of AWS; if the way people use computing shifts, and AWS isn't well-configured for AI workloads, then AWS has a lot to lose.
> Every other player is scrambling for hardware/electricity while Amazon has been building out data centers for the last 20 years.
Microsoft and Google have also been building out data centers for quite a while, but also haven't sat out the AI talent wars the way Amazon has.
1. Price-performance has struggled to stay competitive. There’s some supply-demand forces at play, but the top companies consistently seem to strike better deals elsewhere.
2. The way AWS is architected, especially on networking, isn’t ideal for AI. They’ve dug their heels on in their own networking protocols despite struggling to compete on performance. I personally know of several workloads that left AWS because they couldn’t compete on networking performance.
3. Struggling on the managed side. On paper a service like Bedrock should be great but in practice it’s been a hot mess. I’d love to use Anthropic via Bedrock, but it’s just much more reliable when going direct. AWS has never been great at these sort of managed services at scale and they’re again struggling here.
In theory they should, but it’s increasingly looking like they’re struggling to attract/retain the right talent to take advantage of that position. On paper they should be wiping the floor with others in this space. In practice they’re getting their *ss kicked and in a panic on what to do.
My understanding is that they fell behind on offering the latest gen Nvidia hardware (Blackwell/Blackwell Ultra) due to their focus on internally developed ASICs (Trainium/Inferentia gen 2).
I'd argue that Meta's income derives in no small part from their best in class ad targeting.
Being on the forefront of
(1) creating a personalized, per user data profile for ad-targeting is very much their core business. An LLM can do a very good job of synthesizing all the data they have on someone to try predicting things people will be interested in.
(2) by offering a free "ask me anything" service from meta.ai which is tied directly to their real-world human user account. They gather an even more robust user profile.
This isn't in-my-opinion simply throwing billions at a problem willy nilly. Figuring out how to apply this to their vast reams of existing customer data economically is going to directly impact their bottom line.
5 minutes on facebook being force-fed mesopotamian alien conspiracies is all you'll need to experience to fully understand just how BADLY they need some kind of intelligence for their content/advertising targeting, artificial or not...
You probably don't spend enough time on their sites to have a good ad targeting model of you developed. The closer you are to normal users, with hundreds of hours of usage and many ad clicks, the more accurate the ads will be for you.
Same terrible experience for me while I was on FB.
I was spending a lot of time there and I do shop a lot online. They couldn’t come with relevant ad targeting for me.
For my wife they started to show relevant ads AFTER she went to settings and manually selected areas she is interested in.
This is not an advanced technology everyone claim FB has.
People look at all the chaos in their AI lab but ignore the fact that they yet again beat on earnings substantially and directly cited better ad targeting as the reason for that. Building an LLM is nice for them, but applying AI to their core business is what really matters financially, and that seems like it's going just fine.
The largest LLMs are mostly going to be running in the cloud, so the general purpose cloud providers (Amazon, Microsoft, Google) are presumably going to be in the business of serving models, but that doesn't necessarily mean they need to build the models themselves.
LLMs look to be shaping up as an interchangeable commodity as training datasets, at least for general purpose use, converge to the limits of the available data, so access to customers seems just as important, if not more, than the models themselves. It seems it just takes money to build a SOTA LLM, but the cloud providers have more of a moat, so customer access is perhaps the harder part.
Amazon do of course have a close relationship with Anthropic both for training and serving models, which seems like a natural fit given the whole picture of who's in bed with who, especially as Anthropic and Amazon are both focused on business customers.
Sure, but you can also sell something without having built it yourself, just as Microsoft Copilot supports OpenAI and Anthropic models.
It doesn't have to be either/or of course - a cloud provider may well support a range of models, some developed in house and some not.
Vertical integration - a cloud provider building everything they sell - isn't necessarily the most logical business model. Sometimes it makes more sense to buy from a supplier, giving up a bit of margin, than build yourself.
I'm just an observer. Microsoft has invested billions in OpenAI and can access their IP as a result. It might even be possible MS hopes that OpenAI fails and doesn't allow them to restructure to continue to acquire outside funding. You can go directly to the announcement of their in-house model offerings and they are clearly using this as a recruiting tool for talent. Whether it makes sense for the cloud providers to build their own models is not for me to say, but they may not have a choice given how quickly OpenAI/Anthropic are burning cash. If those two fail then they're essentially ceding the market to Google.
I think this analysis is a bit shallow with regard to Metas product portfolio and how AI fits in.
Much more than the others, metter runs a content business. Gen AI aides in content generation so it behooves them to research it. Even before the current explosion of chatbots, meta was putting this stuff into their VR framework. It's used for their headset tracking and speech to text is helpful for controlling a headset without a physical keyboard.
You're making it sound like they'll follow anything that walks by but I do think it's more strategic than that.
But that didn't require deep insight. Both were already really popular and clearly a threat to Facebook. WhatsApp was huge in Europe before they bought (possibly other places as well).
Buying competition is par for the course for near-monopolies in their niches. As long as the scale differences in value are still very large, you can avoid competition relatively cheaply, while the acquired still walk away with a lot of money.
Why does investing in AI require deep insight? ChatGPT is already huge, significantly bigger than Whatsapp when the deal was done. And while OpenAI is not for sale, he figured that their employees are. Also not to mention, investors are very positive for AI.
So far there hasn't been a transformative use case for LLMs besides the straightforward chat interface (Or some adjacent derivative). Cursor and IDE extensions are nice, but not something that generates billions in revenue.
This means there's two avenues:
1. Get a team of researchers to improve the quality of the models themselves to provide a _better_ chat interface
2. Get a lot of engineers to work LLMs into a useful product besides a chat interface.
I don't think that either of these options are going to pan out. For (1), the consumer market has been saturated. Laymen are already impressed enough by inference quality, there's little ground to be gained here besides a super AGI terminator Jarvis.
I think there's something to be had with agentic interfaces now and in the future, but they would need to have the same punching power to the public that GPT3 did when it came out to justify the billions in expenditure, which I don't think it will.
I think these companies might be able to break even if they can automate enough jobs, but... I'm not so sure.
Whatsapp had $10M revenue when it was acquired[1]. Lots of so called "chatgpt wrappers" has more revenue than that. While in hindsight Whatsapp acquisition at $19B seems no brainer, no concrete metric pointed to that compared to him investing $19B in AI now.
Dude Zuckerberg bought whatsapp because FB Messenger was losing market share... nothing to do with Whatsapps revenue! Rather Zuckerbergs fear of FB products being displaced.
How many software engineers are there in the world? How many are going to stop using it when model providers start increasing token cost on their APIs?
I could see the increased productivity of using Cursor indirectly generating a lot more value per engineer, but... I wouldn't put my money on it being worth it overall, and neither should investors chasing the Nvidia returns bag.
Amazon strategy is to invest in the infrastructure, money is where the machines live. I think they just realized none of those companies have a moat, so why would they? But all of them will buy compute
Except they’re struggling here. The performance of their offerings is consistently behind competitors, particularly given their ongoing networking challenges, and they’re consistently undercut on pricing.
For Amazon “renting servers” at very high margin is their cash cow. For many competitors it’s more of a side business or something they’re willing to just take far lower margin on. Amazon needs to keep the markup high. Take away the AWS cash stream and the whole of Amazon’s financials start to look ugly. That’s likely driving the current panic with its leadership.
Culturally Amazon does really well when it’s an early mover leader in a space. It really struggles, and its leadership can’t navigate, when it’s behind in a sector as is playing out here.
Under what scenario does Amazon lose the beast that is its high margin cloud service renting? It appears to be under approximately zero threat.
Companies are not going to stop needing databases and the 307 other things AWS provides, no matter how good LLMs get.
Cheaper competitors have been trying to undercut AWS since the early days of its public availability, it has not worked to stop them at all. It's their very comprehensive offering, proven track record and the momentum that has shielded AWS and will continue to indefinitely.
It’s already playing out. Just look at recent results. While once light years ahead competitors are now closing ranks and margins are under pressure. AWS clearly isn’t going away, but on the current trajectory its future as the leading cloud is very much not a certainty.
Because if LLM inference is going to be a bigger priority for the majority of companies, they're going to go where they can get the best performance to cost ratio. AWS is falling behind on this. So companies (especially new ones) are going to start using GCP or Azure, and if they're already there for their LLM workloads, why not run the rest of the infrastructure there?
It's similar to how AWS became the de-facto cloud provider for newer companies. They struggled to convince existing Microsoft shops to migrate to AWS, instead most of the companies just migrated to Azure. If LLMs/AI become a major factor in new companies deciding which will be their default cloud provider, they're going to pick GCP or Azure.
Except for spending cloud budgets on LLMs elsewhere like other mentioned, LLM coding will make it easier to convert codebases from being AWS dependent, easing their lock-in
Microsoft has the pleasure of letting you pay for your own hosted GPT models, Mixtral, etc
Microsoft's in a sweet spot. Apple's another interesting one, you can run local LLM models on your Mac really nicely. Are they going to outcompete an Nvidia GPU? Maybe not yet, but they're fast enough as-is.
To do what for that money? Write summaries of product reviews? If they wanted to do something useful, they'd use the LLM to figure out which reviews are for a different product than what is currently being displayed.
> But actually, Amazon, Apple etc aren't natural homes for this, they don't need to burn money to chase it.
I really liked the concept of Apple Intelligence with everything happening all on device, both process and data with minimal reliance off device to deliver the intelligence. It’s been disappointing that it hasn’t come to fruition yet. I am still hopeful the vapor materializes soon. Personally I wouldn’t mind seeing them burning a bit more to make it happen.
It will likely occur, just maybe not this year or next. If we look over the last eighty years of computing, the trend has been smaller and more powerful computers. No reason to think this won’t occur with running inference on larger models.
Exactly. Being a tech company doesn't mean you need to do everything any more than just because you're a family doctor you also should do trauma surgery, dentistry, and botox injections. Pick a lane, be an expert in it.
Except that Amazon's AWS business is severely threatened by the rise of alternative cloud providers who offer much more AI-friendly environments. It's not an existential topic yet, but could easily turn into one.
Zuckerbergs AI "strategy" seems to be to make it easy for people to generate AI slop and share it on FB thus keeping them active on the platform. Or to give people AI "friends" to interact with on FB, thus keeping them on the platform and looking at ads. It's horrifying but it does make business sense (IMHO) at least at first glance.
They have out sensors though for any AGI, because AGI could subvert buisness fields and expertise moats. Thats what most AI teams are- vanity projects and a few experts calming the higher ups every now and then with a "its still just autocompletion on steroids, it can not yet do work and research alone."
> Zuckerberg rushing into every new fad with billions of dollars has somehow tricked people into thinking that's what big tech is about and all of them should be shovelling money into this.
Zuckerberg failed every single fad he tried.
He's becoming more irrelevant every year and only the company's spoils from the past (earned not less by enabling, for example, a genocide to be committed in Myanmar https://www.pbs.org/newshour/world/amnesty-report-finds-face...) help carry them through to the series of disastrous idiotic decision Zuck is inflicting on them.
- VR with Oculus. It never caught on, for most people who own one, it's just gathering dust.
He is doing it at the worst possible moment: LLMs are stagnating and even far better players than Meta like Anthropic and OpenAI can't produce anything worth writing about.
ChatGPT5 was a flop, Anthropic are struggling financially and are lowering token limits and preparing users for cranking up prices, going 180 on their promises not to use chat data for training, and Zuck, in his infinite wisdom, decides to hire top AI talent for premium price at a rapidly cooling market? You can't make up stuff like that.
It would appear that apart from being an ass kisser to Trump, Zuck shares another thing with the orange man-child running the US: a total inability to make good, or even sane deals. Fingers crossed that Meta goes bankrupt just like Trump's 6 banrkruptcies and then Zuck can focus on his MMA career.
I've been taking heat for years for making fun of the metaverse. I had hopeful digital landlords explain to me that theyll be charging rent in there! Who looked at that project and thought it was worth anything?
> I don't know in what circles you're hanging out, I don't know a single person who believed in the metaverse
Oh please, the world was full of hype journalists wanting to sound like they get it and they are in it, whatever next trash Facebook throws their way.
The same way folks nowadays pretend like the LLMs are the next coming of Jesus, it's the same hype as the scrum crowd, the same as crypto, nfts, web3. Always ass kissers who cant think for themselves and have to jump on some bandwagon to feign competence.
meta made $62 billion dollars last year. Mark burns all this money because his one and only priority is making sure his company doesnt become an also ran. The money means nothing to him
Google basically invented modern AI (the 'T' in ChatGPT stands for Transformer), then took a very broad view of how to apply broadly neural AI - AlphaGo, AlphaGenome being the kind of non-LLM stuff they've done).
A better way to look at it is that the absolute number 1 priority for google since they first created a money spiggot throguh monetising high-intent search and got the monopoly on it (outside of Amazon) has been to hold on to that. Even YT (the second biggest search engine on the internet other than google itself) is high intent search leading to advertising sales conversion.
So yes, google has adopted and killed lots of products, but for its big bets (web 2.0 / android / chrome) it's basically done everything it can to ensure it keeps it's insanely high revenue and margin search business going.
What it has to show for it is basically being the only company to have transitioned as dominent across technological eras (desktop -> web2.0 -> mobile -> (maybe llm).
As good as OpenAI is as a standalone, and as good as Claude / Claude Code is for developers, google has over 70% mobile market share with android, nearly 70% browser market share with chrome - this is a huge moat when it comes to integration.
You can also be very bullish about other possible trends. For AI - they are the only big provider which has a persistent hold on user data for training. Yes, OpenAI and Grok have a lot of their own data, but google has ALL gmail, high intent search queries, youtube videos and captions, etc.
And for AR/VR, android is a massive sleeping giant - no one will want to move wholesale into a Meta OS experience, and Apple are increasingly looking like they'll need to rely on google for high performance AI stuff.
All of this protects google's search business a lot.
Don't get me wrong, on the small stuff google is happy to let their people use 10% time to come up with a cool app which they'll kill after a couple of years, but for their big bets, every single time they've gone after something they have a lot to show for it where it counts to them.
Yeah, and Google has cared deeply about AI as a long term play since before they were public. And have been continuously invested there over the long haul.
The small stuff that they kill is just that--small stuff that was never important to them strategically.
I mean, sure, don't heavily invest (your attention, time, business focus, whatever) in something that is likely to be small to Google, unless you want to learn from their prototypes, while they do.
But to pretend that Google isn't capable of sustained intense strategic focus is to ignore what's clearly visible.
I haven't followed that closely, but Gemini seems like a pivot based on ChatGPT's market success
Google is leading in terms of fundamental technology, but not in terms of products
They had the LLambda chatbot before that, but I guess it was being de-emphasized, until ChatGPT came along
Social was a big pivot, though that wasn't really due to Pichai. That was while Larry Page was CEO and he argued for it hard. I can't say anyone could have known beforehand, but in retrospect, Google+ was poorly conceived and executed
---
I also believe the Nth Google chat app was based on WhatsApp success, but I can't remember the name now
Google Compute Engine was also following AWS success, after initially developling Google App Engine
>I haven't followed that closely, but Gemini seems like a pivot based on ChatGPT's market success
"AI" in it's current form is already a massive threat to Google's main business (I personally use Google only a fraction of what I used to), so this pivot is justified.
If you are defining "pivot" as "abandon all other lines of business", then no, none of the BigTechs have ever pivoted.
By more reasonable standards of "pivot", the big investment into Google Plus/Wave in the social media era seems to qualify. As does the billions spent building out Stadia's cloud gaming. Not to mention the billions invested in their abandoned VR efforts, and the ongoing investment into XR...
I'd personally define that as Google hedging their bet's and being prepared in case they needed to truly pivot, and then giving up when it became clear that they wouldn't need to. Sort of like "Apple Intelligence" but committing to the bit, and actually building something that was novel, and useful to some people, who were disappointed when it went away.
Stadia was always clearly unimportant to Google, and I say that as a Stadia owner (who got to play some games, and then got refunds.) As was well reported at the time, closing it was immaterial to their financials. Just because spending hundreds of millions of dollars or even a few billion dollars is significant to you or I doesn't mean that this was ever part of their core business.
Regardless, the overall sentimentality on HN about Google Reader and endless other indisputably small projects says more about the lack of strategic focus from people here, than it says anything about Alphabet.
> Well, "pivot" implies the core business has failed and you're like "oh shit, let's do X instead".
I mean, Facebook's core business hasn't actually failed yet either, but their massive investments in short-form video, VR/XR/Metaverse, blockchain, and AI are all because they see their moat crumbling and are desperately casting around for a new field to dominate.
Google feels pretty similar. They made a very successful gambit into streaming video, another into mobile, and a moderately successful one into cloud compute. Now the last half a dozen gambits have failed, and the end of the road is in sight for search revenue... so one of the next few investments better pay off (or else)
The link you posted has a great many very insignificant investments included in it, and nothing I've seen Google doing has felt quite like the desperation of Facebook in recent years.
I didn't really see it at first, but I think you are correct to point out that they kind of rhyme. However to me, I think the clear desperation of Facebook makes it feel rather different from what I've seen Google doing over the years. I'm not sure I agree that Google's core business is in jeopardy in the way that Facebook's aging social media platform is.
I suppose you could argue that Amazon does have one special thing going for it here, idle compute resources in AWS. However that is not the sort of thing that requires "AI talent" to make use of.
The evidence shows that there is no methodological moat for LLMS. The moat of the frontier folks is just compute. xAI went in months from nothing to competing with the top dogs. DeepSeek too. So why bother with splurging billions in talent when you can buy GPUs and energy instead and serve the compute needs of everyone?
Also Amazon is in another capital intensive business. Retail. Spending billions on dubious AWS moonshots vs just buying more widgets and placing them across the houses of US customers for even faster deliveries does not make sense.
A lot of C-suite people seem to have an idea that if they just throw enough compute at LLMs that AGI will eventually emerge, even though it's pretty clear at this point that LLMs are never going to lead to general intelligence. In their view it makes sense to invest massive amounts of capital because it's like a lottery ticket to being the future AGI company that dominates the world.
I recall Zuckerberg saying something about how there were early signs of AI "improving itself." I don't know what he was talking about but if he really believes that's true and that we're at the bottom of an exponential curve then Meta's rabid hiring and datacenter buildout makes sense.
In early 2023, I remember someone breathlessly explaining that there are signs that LLMs that are seemingly good at chess/checkers moves may have a rudimentary model of the board within them, somehow magically encoded into the model weights through the training. I was stupid enough to briefly entertain the possibility until I actually bothered to develop a high level understanding of the transformer architecture. It's surprising how much mysticism this field seems to attract. Perhaps it being a non-deterministic, linguistically invoked black box, triggers the same internal impulses that draw some people to magic and spellcasting.
Just because it's not that hard to reach a high-level understanding of the transformer pipeline doesn't mean we understand how these systems function, or that there can be no form of world model that they are developing. Recently there has been more evidence for that particular idea [1]. The feats of apparent intelligence LLMs sometimes display have taken even their creators by surprise. Sure, there's a lot of hype too, that's part and parcel of any new technology today, but we are far from understanding what makes them perform so well. In that sense, yeah you could say they are a bit "magical".
> Just because it's not that hard to reach a high-level understanding of the transformer pipeline doesn't mean we understand how these systems function
Mumbo jumbo magical thinking.
They perform so well because they are trained on probabilistic token matching.
Where they perform terribly, e.g math, reasoning, they are delegating to other approaches, and that's how you get the illusion that there is actually something there. But it's not. Faking intelligence is not intelligence. It's just text generation.
> In that sense, yeah you could say they are a bit "magical"
Nobody but the most unhinged hype pushers are calling it "magical". The LLM can never ever be AGI. Guessing the next word is not intelligence.
> there can be no form of world model that they are developing
Kind of impossible to form a world model if your foundation is probabilistic token guessing which is what LLMs are. LLMs are a dead end in achieving "intelligence", something novel as an approach needs to be discovered (or not) to go into the intelligence direction. But hey, at least we can generate text fast now!
There is no evidence to indicate this is the case. To the contrary, all evidence we have points to these models, over time, being able to perform a wider range of tasks at a higher rate of success. Whether it's GPQA, ARC-AGI or tool usage.
> they are delegating to other approaches
> Faking intelligence is not intelligence. It's just text generation.
It seems like you know something about what intelligence actually is that you're not sharing. If it walks, talks and quacks like a duck, I have to assume it's a duck[1]. Though, maybe it quacks a bit weird.
> There is no evidence to indicate this is the case
Burden of proof is on those trying to convince us to buy into the idea of LLMs as being "intelligence".
There is no evidence of the Flying Spaghetti monster or Zeus or God not existing either, but we don't take seriously the people who claim they do exist (and there isn't proof because these concepts are made up).
Why should we take seriously the tolks claiming LLMs are intelligence without proof (there can't be proof, of course, because LLMs are not intelligence)?
Is there something we are all missing? Using Claude feels like magic sometimes, but can't everyone see the limitation now that we are 4 years and 100s of billions down the road?
Are they still really hoping that they are gonna tweak a model and feed it an even bigger dataset and it will be AGI?
I'm not a fan of mysticism. I'm also with you that these are simply statistical machines. But I don't understand what happened when understood transformers at a high-level.
If you're saying the magic disappeared after looking at a single transformer, did the magic of human intelligence disappear after you understood human neurons at a high level?
Tbh I find this view odd, and I wonder what people view as agi now. It used to be that we had extremely narrow pieces of AI and I remember being on a research project about architectures and just very basic “what’s going on?” was advanced. Understanding that someone asked a question, that would be solved by getting a book and being able to then go and navigate to the place the book was likely to be was fancy. Most systems could solve literally one type of problem. They weren’t just bad at other things they were fundamentally incapable of anything but an extremely narrow use case.
I can throw wide ranging problems at things like gpt5 and get what seem like dramatically better answers than if I asked a random person. The amount of common sense is so far beyond what we had it’s hard to express. It used to be always pointed out that the things we had were below basic insect level. Now I have something that can research a charity, find grants and make coherent arguments for them, read matrix specs and debug error messages, and understand sarcasm.
To me, it’s clear that agi is here. But then what I always pictured from it may be very different to you. What’s your image of it?
It's more that "random" people are dumb as bricks (but we've in the name of equality and historic measurement errors decided to forgo that), add to it that AI's have a phenomenal (internet sized) memory makes them far more capable than many people.
However, even "dumb" people can often make judgements structures in a way that AI's cannot, it's just that many have such a bad knowledge-base that they cannot build the structures coherently whereas AI's succeed thanks to their knowledge.
I wouldn't be surprised if the top AI firms today spend an inordinate amount of time to build "manual" appendages into the LLM systems to cater to tasks such as debugging to uphold the facade that the system is really smart, while in reality it's mostly papering up a leaky model to avoid losing the enormous investments they need to stay alive with a hope that someone on their staff comes up a real solution to self-learning.
I think the discrepancy between different views on the matter mainly stems from the fact that state-of-the-art LLMs are better (sometimes extremely better) at some tasks, and worse (sometimes extremely worse) at other tasks, compared to average humans. For example, they're better at retrieving information from huge amounts of unstructured data. But they're also terrible at learning: any "experience" which falls out of the context window is lost forever, and the model can't learn from its mistakes. To actually make it learn something requires very many examples and a lot of compute, whereas a human can permanently learn from a single example.
You discover truth by doing stuff in real world and observing the results. Current LLM have enough intelligence, but all the inputs they have are the “he said she said” by us monkeys, including all omissions and biases.
It's not about what the average human can do - it's about what humans as a category are capable of. There will always be outliers (in both directions), but you can, in general, teach a human a variety of tasks, such as performing arithmetic deterministically, that you cannot teach to, for example, a parrot.
I don’t have a very high expectation of AGI at all. Just an algorithm or system you can put onto a robot dog, and get a dog level general intelligence. You should be able to live with that robot dog for 10 years and it should be just as capable as a dog throughout that timespan.
Hell, I’d even say we have AGI if you could emulate something like a hamster.
LLMs are way more impressive in certain ways than such a hypothetical AGI. But that has been true of computers for a long time. Computers have been much better at Chess than humans for decades. Dogs can’t do that. But that doesn’t mean that a chess engine is an AGI.
I would also say we have a special form of AGI if the AI can pass an extended Turing test. We’ve had chat bots that can fool a human for a minute for a long time. Doesn’t mean we had AGI. So time and knowledge was always a factor in a realistic Turing test. If an AGI can fool someone who knows how to properly probe an LLM, for a month or so, while solving a bunch of different real world tasks that require stable long term memory and planning, then I’d day we’re in AGI territory for language specifically. I think we have to distinguish between language AGI and multi-modal AGI. So this test wouldn’t prove what we could call “full” AGI.
These are some of the missing components for full AGI:
- Being able to act as a stable agent with a stable personality over long timespans
- Capable of dealing with uncertainties. Having a understanding of what it doesn’t know
- One-shot learning, with long term retention, for a large number of things
- Fully integrated multi-modality across sound, vision, and other inputs/outputs we may throw at it.
The last one is where we may be able to get at the root of the algorithm we’re missing. A blind person can learn to “see” by making clicks and using their ears to see. Animals can do similar “tricks”. I think this is where we truly see the full extent of the generality and adaptability of the biological brain. Imagine trying to make a robot that can exhibit this kind of adaptability. It doesn’t fit into the model we have for AI right now.
I wonder if 2) is a result of published bias for positive results in the training set. An “I don’t know” response is probably ranked unsatisfactory by human feedback and most published scientific literature are biased towards positive results and factual explanations.
In my experience, the willingness to say "I don't know" instead of confabulate is also down-rated as a human attribute, so it's not surprising that even an AGI trained on the "best" of humanity would avoid it.
Recently I realized that US are very close to a centrally planned economy. Meta wasted 50B on metaverse, which like how much Texas spends on healthcare. Now the "AI" investments seems dubious.
You could fund 1000+ projects with this kinds of money. This is not an effective capital allocation.
> ... and has an idea of how they work shouldn't think its going to lead to "AGI"
Not sure what level of understanding are you referring to but having learned and researched about the pretty much all LLM internals I think this has led me exactly to the opposite line of thinking. To me it's unbelievable what we have today.
I think AI research is like anything else really. The smartest people are heads down working on their problems. The people going on podcasts are less connected to day to day work.
It’s also pretty useless to talk about whether something is AGI without defining intelligence in the first place.
I think something like we saw in the show "Devs" is much more likely, although what the developers did with it in the show was bonkers unrealistic. But some kind of big enough quantum device basically.
This is not and has not been the consensus opinion. If you're not an AI researcher you shouldn't write as if you've set your confidence parameter to 0.95.
Of course it might be the case, but it's not a thing that should be expressed with such confidence.
Is it widely accepted that LLMs won't lead to AGI? I've asked Gemini, so it came up with four primary arguments for this claim, commenting on them briefly:
1) LLMs as simple "next token predictors" so they just mimicry thinking: But can it be argued that current models operate on layers of multiple depth and are able to actually understand by building concepts and making connections on abstract levels? Also, don't we all mimicry?
2) Grounding problem: Yes, models build their world models on text data, but we have models operating on non-textual data already, so this appears to be a technical obstacle rather than fundamental.
3) Lack of World Model. But can anyone really claim they have a coherent model of reality? There are flat-earthers, yet I still wouldn't deny them having AGI. People hallucinate and make mistakes all the time. I'd argue hallucinations is in fact the sign of an emerging intelligence.
4) Fixed learning data sets. Looks like this is now being actively solved with self-improving models?
I just couldn't find a strong argument supporting this claim. What am I missing?
Why on earth would you copy and paste an LLM's output into a comment? What does that accomplish or provide that just a simply stated argument doesn't accomplish more succinctly? If you don't know something, simply don't comment on it -- or ask a question.
> I've asked Gemini, so it came up with four primary arguments for this claim, commenting on them briefly:
This line means, and literally says, that everything that follows is a summary or direct quotation from an LLM's output.
There's a more charitable but unintuitive interpretation, in which "commenting on them briefly" is intended to mean "I will comment on them briefly:". But this isn't a natural interpretation. It's one I could be expected to reach only after seeing your statement that 'none of the above is AI.' But even this more charitable interpretation actually contradicts your claim that it's not AI.
So now I'm even less sure I know what you meant to communicate. Either I'm missing something really obvious or the writing doesn't communicate what you intended.
Fur future reference, pasting llm slop feels exactly as patronizing as back when people pasted links to google searches in response to questions they considered beneath their dignity to answer. Except in this case, no-one asked to begin with.
There's a bunch of ways AI is improving itself, depending on how you want to interpret that. But it's been true since the start.
1. AI is used to train AI. RLHF uses this, curriculum learning is full of it, video model training pipelines are overflowing with it. AI gets used in pipelines to clean and upgrade training data a lot.
2. There are experimental AI agents that can patch their own code and explore a tree of possibilities to boost their own performance. However, at the moment they tap out after getting about as good as open source agents, but before they're as good as proprietary agents. There isn't exponential growth. There might be if you throw enough compute at it, but this tactic is very compute hungry. At current prices it's cheaper to pay an AI expert to implement your agent than use this.
So have an AI with a 40% error rate judge an AI with an 40% error rate…
AGI is a complete no go until a model can adjust its own weights on the fly, which requires some kind of negative feedback loop, which requires a means to determine a failure.
Humans have pain receptors to provide negative feedback and we can imagine events that would be painful such as driving into a parked car would be painful without having to experience it.
If current models could adjust its own weights to fix the famous “how many r’s in strawberry” then I would say we are on the right path.
However, the current solution is to detect the question and forward it to a function to determine the right answer. Or attempt to add more training data the next time the model is generated ($$$). Aka cheat the test.
I think LLM as a toolsmith like demonstrated in the Voyager paper (1) is another interesting approach to creating a system that can learn to do a task better over time. (1) https://arxiv.org/abs/2305.16291
> There are experimental AI agents that can patch their own code and explore a tree of possibilities to boost their own performance. However, at the moment they tap out after getting about as good as open source agents, but before they're as good as proprietary agents.
I'm skeptical that RLHF really works. Doesn't it just patch the obvious holes so it looks better on paper? If it can't reason then it will continue to get 2nd and 3rd order difficulty problems wrong.
> it's pretty clear at this point that LLMs are never going to lead to general intelligence.
It is far from clear. There may well be emergent hierarchies of more abstract thought at much higher numbers of weights. We just don't know how a transformer will behave if one is built with 100T connections - something that would finally approach the connectome level of a human brain. Perhaps nothing interesting but we just do not know this and the current limitation in building such a beast is likely not software but hardware. At these scales the use of silicon transistors to approximate analog curve switching models just doesn't make sense. True neuromorphic chips may be needed to approach the numbers of weights necessary for general intelligence to emerge. I don't think there is anything in production at the moment that could rival the efficiency of biological neurons. Most likely we do not need that level of efficiency. But it's almost certain that stringing together a bunch of H100s isn't a path to the scale we should be aiming for.
Even assuming a company gets to AGI first this doesn't mean another one will follow.
Suppose that FooAI gets to it first:
- competitors may get there too in a different or more efficient way
- Some FooAI staff can leave and found their own company
- Some FooAI staff can join a competitor
- FooAI "secret sauce" can be figured out, or simply stolen, by a competitor
At the end of the day, it really doesn't matter, the equation AI === commodity just does not change.
There is no way to make money by going into this never ending frontier model war, price of training keeps getting higher and higher, but your competitors few months later can achieve your own results for a fraction of your $.
Some would say that the race to AGI is like the race to nuclear weapons and that the first to get there will hold all the cards (and be potentially able to stop others getting there.) It's a bit too sci-fi for me.
If AGI is reached it would be trivial for the competing superpowers to completely quarantine themselves network wise by cutting undersea cables long enough to develop competing AGI
I don't know if AGI will emerge from LLM, but I'm always reminded of the Chinese room thought experiment. With billions thrown at the idea it will certainly be the ultimate answer as to whether true understanding can emerge from a large enough dictionary.
Please stop refering to the Chinese Room - it's just magical/deist thinking in disguise. It postulates that humans have way of 'understanding' things that is impossible to replicate mechanically.
The fact that philosophy hasn't recognized and rejected this argument based on this speaks volumes of the quality of arguments accepted there.
(That doesn't mean LLMs are or will be AGI, its just this argument is tautological and meaningless)
That some people use the Chinese Room to ascribe some magical properties to human consciousness says more about the person drawing that conclusion than the thought experiment itself.
I think it's entirely valid to question whether a computer can form an understanding through deterministically processing instructions, whether that be through programming language or language training data.
If the answer is no, that shouldn't lead to a deist conclusion. It can just as easily lead to the conclusion that a non-deterministic Turing machine is required.
I'd appreciate if you tried to explain why instead of resorting to ad hominem.
> I think it's entirely valid to question whether a computer can form an understanding through deterministically processing instructions, whether that be through programming language or language training data.
Since the real world (including probabilistic and quantum phenomena) can be modeled with deterministic computation (a pseudorandom sequence is deterministic, yet simulates randomness), if we have a powerful enough computer we can simulate the brain to a sufficient degree to have it behave identically as the real thing.
The original 'Chinese Room' experiment describes a book of static rules of Chinese - which is probably not the way to go, and AI does not work like that. It's probabilistic in its training and evaluation.
What you are arguing is that constructing an artificial consciousness lies beyond our current computational ability(probably), and understanding of physics (possibly), but that does not rule out that we might solve these issues at some point, and there's no fundamental roadblock to artificial consciousness.
I've re-read the argument (https://en.wikipedia.org/wiki/Chinese_room) and I cannot help but conclude that Searle argues that 'understanding' is only something that humans can do, which means that real humans are special in some way a simulation of human-shaped atoms are not.
Which is an argument for the existence of the supernatural and deist thinking.
> I'd appreciate if you tried to explain why instead of resorting to ad hominem.
It is not meant as an ad hominem. If someone thinks our current computers can't emulate human thinking and draws the conclusion that therefore humans have special powers given to them by a deity, then that probably means that person is quite religious.
I'm not saying you personally believe that and therefore your arguments are invalid.
> Since the real world (including probabilistic and quantum phenomena) can be modeled with deterministic computation (a pseudorandom sequence is deterministic, yet simulates randomness), if we have a powerful enough computer we can simulate the brain to a sufficient degree to have it behave identically as the real thing.
The idea that a sufficiently complex pseudo-random number generator can emulate real-world non-determinism enough to fully simulate the human brain is quite an assumption. It could be true, but it's not something I would accept as a matter of fact.
> I've re-read the argument (https://en.wikipedia.org/wiki/Chinese_room) and I cannot help but conclude that Searle argues that 'understanding' is only something that humans can do, which means that real humans are special in some way a simulation of human-shaped atoms are not.
In that same Wikipedia article Searle denies he's arguing for that. And even if he did secretly believe that, it doesn't really matter, because we can draw our own conclusions.
Disregarding his arguments because you feel he holds a hidden agenda, isn't that itself an ad hominem?
(Also, I apologize for using two accounts, I'm not attempting to sock puppet)
>Searle argues that, without "understanding" (or "intentionality"), we cannot describe what the machine is doing as "thinking" and, since it does not think, it does not have a "mind" in the normal sense of the word.
This is the only sentence that seems to be pointing to what constitutes the specialness of humans, and the terms of 'understanding' and 'intentionality' are in air quotes so who knows? This sounds like the archetypical no true scotsman fallacy.
In mathematical analysis, if we conclude that the difference between 2 numbers is smaller than any arbitrary number we can pick, those 2 numbers must be the same. In engineering, we can reduce the claim to 'any difference large about to care about'
Likewise if the difference between a real human brain and an arbitrarily sophisticated Chinese Room brain is arbitrarily small, they are the same.
If our limited understanding of physics and engineering makes the practical difference not zero, this essentially becomes a bit of a somewhat magical 'superscience' argument claiming we can't simulate the real world to a good enough resolution that the meaningful differences between our 'consciousness simulator' and the thing itself disappear - which is an extraordinary claim.
They're in the "Complete Argument" section of the article.
> This sounds like the archetypical no true scotsman fallacy.
I get what you're trying to say, but he is not arguing only a true Scotsman is capable of thought. He is arguing that our current machines lack the required "causal powers" for thought. Powers that he doesn't prescribe to only a true Scotsman, though maybe we should try adding bagpipes to our AI just to be sure...
Thanks, but that makes his arguments even less valid.
He argues that computer programs only manipulate symbols and thus have no semantic understanding.
But that's not true - many programs, like compilers that existed back when the argument was made, had semantic understanding of the code (in a limited way, but they did have some understanding about what the program did).
LLMs in contrast have a very rich semantic understanding of the text they parse - their tensor representations encode a lot about each token, or you can just ask them about anything - they might not be human level at reading subtext, but they're not horrible either.
Now you're getting to the heart of the thought experiment. Because does it really understand the code or subtext, or is it just really good at fooling us that it does?
When it makes a mistake, did it just have a too limited understanding or did it simply not get lucky with its prediction of the next word? Is there even a difference between the two?
I would like to agree with you that there's no special "causal power" that Turing machines can't emulate. But I remain skeptical, not out of chauvinism, but out of caution. Because I think it's dangerous to assume an AI understands a problem simply because it said the right words.
> I cannot help but conclude that Searle argues that ‘understanding’ is only something that humans can do, which means…
Regardless of whether Searle is right or wrong, you’ve jumped to conclusions and are misunderstanding his argument and making further assumptions based on your misunderstanding. Your argument is also ad-hominem by accusing people of believing things they don’t believe. Maybe it would be prudent to read some of the good critiques of Searle before trying to litigate it rapidly and sloppily on HN.
The randomness stuff is very straw man, definitely not a good argument, best to drop it. Today’s LLMs are deterministic, not random. Pseudorandom sequences come in different varieties, but they model some properties of randomness, not all of them. The functioning of today’s neural networks, both training and inference, is exactly a book of static rules, despite their use of pseudorandom sequences.
In case you missed it in the WP article, most of the field of cognitive science thinks Searle is wrong. However, they’re largely not critiquing him for using metaphysics, because that’s not his argument. He’s arguing that biology has mechanisms that binary electronic circuitry doesn’t; not human brains, simply physical chemical and biological processes. That much is certainly true. Whether there’s a difference in theory is unproven. But today currently there absolutely is a difference in practice, nobody has ever simulated the real world or a human brain using deterministic computation.
If scientific consensus is that he's wrong why is he being constantly brought up and defended - am I not right to call them out then?
Nobody brings up that light travels through the aether, that diseases are caused by bad humors etc. - is it not right to call out people for stating theory that's believed to be false?
>The randomness stuff is very straw man,
And a direct response to what armada651 wrote:
>I think it's entirely valid to question whether a computer can form an understanding through deterministically processing instructions, whether that be through programming language or language training data.
> He’s arguing that biology has mechanisms that binary electronic circuitry doesn’t; not human brains, simply physical chemical and biological processes.
Once again the argument here changed from 'computers which only manipulate symbols cannot create consciousness' to 'we don't have the algorithm for consiousness yet'.
He might have successfully argued against the expert systems of his time - and true, mechanistic attempts at language translation have largely failed - but that doesn't extend to modern LLMs (and pre LLM AI) or even statistical methods.
You’re making more assumptions. There’s no “scientific consensus” that he’s wrong, there are just opinions. Unlike the straw man examples you bring up, nobody has proven the claims you’re making. If they had, then the argument would go away like the others you mentioned.
Where did the argument change? Searle’s argument that you quoted is not arguing that we don’t have the algorithm yet. He’s arguing that the algorithm doesn’t run on electrical computers.
I’m not defending his argument, just pointing out that yours isn’t compelling because you don't seem to fully understand his, at least your restatement of it isn’t a good faith interpretation. Make his argument the strongest possible argument, and then show why it doesn’t work.
IMO modern LLMs don’t prove anything here. They don’t understand anything. LLMs aren’t evidence that computers can successfully think, they only prove that humans are prone to either anthropomorphic hyperbole, or to gullibility. That doesn’t mean computers can’t think, but I don’t think we’ve seen it yet, and I’m certainly not alone there.
And? Are you imagining that these aren’t both true at the same time? If so, I’m happy to explain. Since nothing has been proven, there’s nothing “scientific”. And since there’s some disagreement, “consensus” has not been achieved yet. This is why your presumptive use of “scientific consensus” was not correct, and why the term “scientific consensus” is not the same thing as “most people think”. A split of 60/40 or 75/25 or even 90/10 counts as “most” but does not count as “consensus”. So I guess maybe be careful about assuming what something means, it seems like this thread was limited by several incorrect assumptions.
> The fact that philosophy hasn't recognized and rejected this argument based on this speaks volumes of the quality of arguments accepted there.
That's one possibility. The other is that your pomposity and dismissiveness towards the entire field of philosophy speaks volumes on how little you know about either philosophical arguments in general or this philosophical argument in particular.
Another ad hominem, I'd like you to refute my claim that Searle's argument is essentially 100% magical thinking.
And yes, if for example, medicine would be no worse at curing cancer than it is today, yet doctors asserted that crystal healing is a serious study, that would reflect badly on the field at large, despite most of it being sound.
Searle refutes your claim that there’s magical thinking.
“Searle does not disagree with the notion that machines can have consciousness and understanding, because, as he writes, "we are precisely such machines". Searle holds that the brain is, in fact, a machine, but that the brain gives rise to consciousness and understanding using specific machinery.”
But the core of the original argument is that programs only manipulate symbols and consciousness can never arise just through symbol manipulation - which here then becomes 'we have not discovered the algorithms' for consciousness yet.
When you say something that contradicts his statements, it doesn’t mean he’s wrong, it most likely means you haven’t understood or interpreted his argument correctly. The Wikipedia page you linked to doesn’t use the word “algorithm”, so the source of the contradiction you imagine might be you. Searle says he thinks humans are biological machines, so your argument should withstand that hypothesis rather than dismiss it.
The human way of understanding things can be replicated mechanically, because it is mechanical in nature. The contents of your skull are an existence proof of AGI.
The contents of my skull are only a proof for AGI if your mechanical machine replicates all its processes. It's not a question about whether a machine can reproduce that, it's a question about whether we have given our current machines all the tools it needs to do that.
The theory of special relativity does not say 'you can't exceed the speed of light(unless you have a really big rocket)'. It presents a theoretical limit. Likewise the Chinese room doesn't state that consciousness is an intractable engineering problem, but an impossibility.
But the way Searle formulates his argument, by not defining what consciousness is, he essentially gives himself enough wiggle room to be always right - he's essentially making the 'No True Scotsman' fallacy.
The moat is people, data, and compute in that order.
It’s not just compute. That has mostly plateaued. What matters now is quality of data and what type of experiments to run, which environments to build.
This "moat" is actually constantly shifting (which is why it isn't really a moat to begin with). Originally, it was all about quality data sources. But that saturated quite some time ago (at least for text). Before RLHF/RLAIF it was primarily a race who could throw more compute at a model and train longer on the same data. Then it was who could come up with the best RL approach. Now we're back to who can throw more compute at it since everyone is once again doing pretty much the same thing. With reasoning we now also opened a second avenue where it's all about who can throw more compute at it during runtime and not just while training. So in the end, it's mostly about compute. The last years have taught us that any significant algorithmic improvement will soon permeate across the entire field, no matter who originally invented it. So people are important for finding this stuff, but not for making the most of it. On top of that, I think we are very close to the point where LLMs can compete with humans on their own algorithmic development. Then it will be even more about who can spend more compute, because there will be tons of ideas to evaluate.
You usually see this from startup techbro CEOs understand neither x nor AI. Those people are already replacable by AI today. The kind of people who think they can query ChatGPT once with "How to create a cutting edge model" and make millions. But when you go in on the deep end, there are very few people who still have enough tech knowledge to compete with your average modern LLM. And even the Math Olympiad gold medalists high-flyers at DeepSeek are about to have a run for their money with the next generation. Current AI engineers will shift more and more towards senior architecture and PM roles, because those will be the only ones that matter. But PM and architecture is already something that you could replace today.
> Originally, it was all about quality data sources.
It still is! Lots of vertical productivity data that would be expensive to acquire manually via humans will be captured by building vertical AI products. Think lawyers, doctors, engineers.
Lets not pretend this is strategy. Amazon has been trying and failing to hire top AI people. No-one in their right minds would join. Even Meta has to shell out 8-9 figures for top people, who with any modicum of talent or self respect would go to Amazon rather than Anthropic, OAI, GDM? They bought Adept, everyone left.
AWS is also falling far behind Azure wrt serving AI needs at the frontier. GCP is also growing at a faster rate and has a way more promising future than AWS in this space.
AWS is very far behind, its already impacting the stock. Without a winning AI offering, all new cloud money is going to GCP and Azure. They have a huge problem
> The evidence shows that there is no methodological moat for LLMS.
Does it? Then how come Meta hasn't been able to release a SOTA model? It's not for a lack of trying. Or compute. And it's not like DeepSeek had access to vastly more compute than other Chinese AI companies. Alibaba and Baidu have been working on AI for a long time and have way more money and compute, but they haven't been able to do what DeepSeek did.
They may not have been leading (as in, releasing a SOTA model), but they definitely can match others - easily, as shown by llama 3/4, which proves the point - there is no moat. With enough money and resources, you can match others. Whether without SOTA models you can make a business out of it is a different question.
> but they definitely can match others - easily, as shown by llama 3/4
Are we living in the same universe? LLAMA is universally recognized as one of the worst and least successful model releases. I am almost certain you haven't ever tried a LLAMA chat, because, by the beard of Thor, it's the worst experience anyone could ever had, with any LLM.
LLAMA 4 (behemoth, whatever, whatever) is an absolute steaming pile of trash, not even close to ChatGPT 4o/4/5/, Gemini(any) and even not even close to cheaper ones like DeepSeek. And to think Meta pirated torrents to train it...
What a bunch of criminal losers and what a bunch of waste of money, time and compute. Oh, at least the Metaverse is a success...
I switched to Gemini with my new phone and I literally couldn't tell a difference. It is actually crazy how small the cost of switching is for LLMs. It feels like AI is more like a commodity than a service.
> I switched to Gemini with my new phone and I literally couldn't tell a difference. It is actually crazy how small the cost of switching is for LLMs. It feels like AI is more like a commodity than a service.
It is. It's wild to me that all these VCs pouring money into AI companies don't know what a value-chain is.
Tokens are the bottom of the value-chain; it's where the lowest margins exist because the product at that level is a widely available commodity.
On top of that, the on-device models have got stronger and stronger as the base models + RL has got better. You can do on your laptop now what 2 years ago was state of the art.
Which dimensions do you see Google lagging on? They seem broadly comparable on the usual leaderboard (https://lmarena.ai/leaderboard) and anecdotally I can't tell the difference in quality.
I tend personally to stick with ChatGPT most of the time, but only because I prefer the "tone" of the thing somehow. If you forced me to move to Gemini tomorrow I wouldn't be particularly upset.
> Which dimensions do you see Google lagging on? They seem broadly comparable on the usual leaderboard (https://lmarena.ai/leaderboard) and anecdotally I can't tell the difference in quality.
Gemini holds indeed the top spot, but I feel you framed your response quite well: they are all broadly comparable. The difference in the synthetic benchmark from the top spot and the 20th spot was something like 57 points on a scale of 0-1500
" in many dimensions they lag behind GPT-5 class " - such as?
Outside of computer, "the moat" is also data to train on. That's an even wider moat. Now, google has all the data. Data no one else has or ever will have. If anything, I'd expect them to outclass everyone by a fat margin. I think we're seeing that on video however.
Do you want to model the world accurately or not? That person is part of our authentic reality. The most sophisticated AI in the world will always include that person(s).
a bit weird to think about it since google has literally internet.zip in multiple versions over the years, all of email, all of usenet, all of the videos, all of the music, all of the user's search interest, ads, everything..
> a bit weird to think about it since google has literally internet.zip in multiple versions over the years, all of email, all of usenet, all of the videos, all of the music, all of the user's search interest, ads, everything..
Yeah, Google totally has a moat. Them saying that they have no moat doesn't magically make that moat go away.
They also own the entire vertical which none of the competitors do - all their competitors have to buy compute from someone who makes a profit just on compute (Nvidia, for example). Google owns the entire vertical, from silicon to end-user.
Given Apple’s moat is their devices, their particular spin on AI is very much edge focussed, which isn’t as spectacular as the current wave of cloud based LLM. Apple’s cloud stuff is laughably poor.
Depending on how you look at it I suppose but I believe Gemini surpasses OpenAI on many levels now. Better photo and video models. The leaderboard for text and embeddings are also putting Google on top of Openai.
gemini-2.5-pro is ranked number 1 in llmarena (https://lmarena.ai/leaderboard) before gpt-5-high. In the Text-to-Video and Image-to-video, google also have the highest places, OpenAI is nowhere.
Yes, but they're also slower. As LLMs start to be used for more general purpose things, they are becoming a productivity bottle-neck. If I get a mostly right answer in a few seconds that's much better than a perfect answer in 5 minutes.
Right now the delay for Google's AI coding assistant is high enough for humans to context switch and do something else while waiting. Particularly since one of the main features of AI code assistants is rapid iteration.
The barriers to entry for LLM are obvious: as you pointed, the field is extremely capital intensive. The only reason there are seemingly multiple players is because the amount of capital thrown at it at the moment is tremendous but that's unlikely to last forever.
From my admittely poorly informed point of view, strategy-wise, it's hard to tell how wise it is investing in foundational work at the moment. As long as some players release competitive open weight models, the competitive advantage of being a leader in R&D will be limited.
Amazon already has the compute power to place itself as a reseller without investing or having to share the revenue generated. Sure, they won't be at the forefront but they can still get their slice of the pie without exposing themselves too much to an eventual downturn.
Think well: why should a platform provider get into a terribly expensive and unprofitable business when they can just provide hardware for those with money to spend? This was AWS strategy for years and it's been working well for them.
Does Amazon want to be an AI innovator or an AI enabler?
AWS enables thousands of other companies to run their business. Amazon has designed their own Graviton ARM CPUS and their own Trainium AI chips. You can access these through AWS for your business.
I think Amazon sees AI being used in AWS as a bigger money generator than designing new AI algorithms.
Also I think that they realize this is just a money losing proposition right now for the most part. And they're not going to have a problem getting in later when there's a clear solution. Why fight it out? I don't think they're going to miss much because they can use any models they need and as you said some of that stuff may be run on their servers
I can make a case: Building their own models like Nova and Titan allow them to build up expertise in how to solve hyperscaler problems. Think of it like Aurora, where they have a generally solved problem (RDBMS) but it needs to be modified to work with the existing low-level primitives. Yes, it can be done in the open, but if I'm AWS, I probably want to jealously guard anything that could be a key differentiator.
Reading comments from the appropriate VPs will illuminate the situation.. Swami is looking to democratise AI, and the company is geared towards that more than anything else.
I started this part of the thread and mentioned Trainium but the person you replied to gave a link. Follow that and you can see Amazon's chips that they designed.
Amazon wants people to move away from Nvidia GPUs and to their own custom chips.
TBH I was just going off of that I've heard AWS is a terrible place to get h100 clusters at scale. And for the training I was looking at we didn't really want to consider going off CUDA.
They have to use GCP as well, which is arguably a strong indictment of their experience with AWS. Coincidentally, this aligns with my experience trying to train on AWS.
It’s unclear why Swami is put in charge of this stuff. He’s not a recognized leader in the space and hasn’t delivered a coherent strategy. However, per the article Amazon is struggling to hire and retain the best talent and thus it may just be the best they have.
I'm glad this analogy is at the top. I think that some large companies like AWS really should not try to blow money on AI in ways that only make a lot more sense for companies like Meta, Google, and Apple. AWS can't trap you in their AI systems with network effects that the other competitors can.
Companies like OpenAI and Anthropic are still incredibly risky investments especially because of the wild capital investments and complete lack of moat.
At least when Facebook was making OpenAI's revenue numbers off of 2 billion active users it was trapping people in a social network where there were real negative consequences to leaving. In the world of open source chatbots and VSClone forks there's zero friction to moving on to some other solution.
OpenAI is making $12 billion a year off of 700 million users [1], or around $17 per user annually. What other products that have no ad support perform that badly? And that's a company that is signing enterprise contracts with companies like Apple, not just some Spotify-like consumer service.
[1] This is almost the exact same user count that Facebook had when it turned its first profit.
> OpenAI is making $12 billion a year off of 700 million users [1], or around $17 per user annually. What other products that have no ad support perform that badly?
That's a bit of a strange spin. Their ARPU is low because they are choosing not to monetize 95% of their users at all, and for now are just providing practically limitless free service.
But monetising those free users via ads will pretty obviously be both practical and lucrative.
And even if there is no technical moat, they seem to have a very solid mind share moat for consumer apps. It isn't enough for competitors to just catch up. They need to be significantly better to shift consumer habits.
(For APIs, I agree there is no moat. Switching is just so easy.)
> They need to be significantly better to shift consumer habits.
i am hoping that a device local model would eventually be possible (may be a beefy home setup, and then an app that connects to your home on mobile devices for use on the go).
currently, hardware restrictions prevent this type of home setup (not to mention the open source/free models aren't quite there and difficulty for non-tech users to actually setup). However, i choose to believe the hardware issues will get solved, and it will merely be just time.
The software/model issue, on the other hand is harder to see solved. I pin my hopes onto deepseek, but may be meta or some other company will surprise me.
i dont have the hardware to run or try them, but from the huggingfaces discussion forums, gpt-oss seems to be pretty hard censored. I would not consider it as being a viable self-hosted LLM except for the very narrowest of domains (like coding for example).
I'm not sure where censorship comes in with this discussion, it seems like cloud models are censored as well? And local models are frequently created that are abliterated? Correct me if I'm wrong or misunderstanding you.
Either way, it's just an example model, plenty of others to choose from. The fact of the matter is that the base model MacBook Air currently comes with about half as much RAM as you need for a really really decent LLM model. The integrated graphics are fast/efficient and the RAM is fast. The AMD Ryzen platform is similarly well-suited.
(Apple actually tells you how much storage their local model takes up in the settings > general > storage if you're curious)
We can imagine that by 2030 your base model Grandma computer on sale in stores will have at least 32GB of high-bandwidth RAM to handle local AI workflows.
which is why i made the claim that hardware "problem" will be solved in the near future (i don't consider it solved right now, because even the apple hardware is too expensive and insufficient imho), but the more difficult problem of model availability is much, much harder to solve.
There does seem to be a mind share mote, but all you have to do is piss off users a little bit when there's a good competitor. See Digg to Reddit exodus.
This is not mutually exclusive. They have home made robots and let others sell robots on their website. The same way they want to use AI and have resources to make their own. One way to use is to drive those robots. Another to enhance their web site. Current version sucks. I recently return the item because their bot told it has functionality while in fact it didn't.
> Of course, the AI talent war may end up being an expensive and misguided strategy, stoked by hype and investor over-exuberance.
To me, that's a pretty good explanation.
The world is crazy with AI right now, but when we see how DeepSeek became a major player at a fraction of the cost, and, according to Google researchers, without making theoretical breakthroughs. It looks foolish to be in this race, especially now that we are seeing diminishing returns. Waiting until things settle, learning from others attempts and designing your system not for top performance but for efficiency and profit seems like a sane strategy.
And it is not like Amazon is out of the AI game, they have what really matters: GPUs. This is a gold rush, and as the saying goes, they are more interested in selling pickaxes that finding gold.
I guess Amazon can also probably afford to wait until somebody comes up with an application for AI that is, like, something Amazon can actually sell or use…
Customer service bots? Maybe. Coding bots? I bet they use some internally. Their customers don’t really need them, or if the customer does, the customer can run it on their side.
As I’ve said before, the kind of AI that makes money is called machine learning. Pricing ads, recommending products, improving search, optimizing routing.
In general these fall into the category of things humans cannot do at the scale and speed necessary to run SaaS companies.
Many of the things LLMs attempt to do are things people already do, slowly and relatively accurately. But until hallucinations are rare, slow expensive humans will typically need to be around. The AI booster’s strategy of ignoring/minimizing hallucinations or equivocating with human fallibility doesn’t work for businesses where reliability is important.
Note that ML algorithms are highly imperfect as well. Uber’s prices aren’t optimal. Google search surfaces tons of spam. But they are better than the baseline of no service exists.
AI is huge, it’s just not the only thing happening in tech right now. I say this as an MLE but it seems really unbalanced that LLMs have gotten trillions in investment when other groundbreaking innovations like battery improvements or fusion power or gene therapy have gotten substantially less attention.
Disagree re: DeepSeek theoretical breakthroughs, MLA and GRPO are pretty good and paved the way for others e.g. Kimi K2 uses MLA for a 1T MoE.
Big money investors know that real tangible products that have real tangible benefits aren't usually decimal-point-shifting-your-net-worth jackpots. They make money, sure, but factories can't be built in a day. Also, if they can make AI work as it says on the box, they'll be able to get rid of all those pesky employees and turn their companies into pure money-printing enterprises.
Pay no attention to the cracks that are showing. Nevermind the chill. Everything is fine.
AWS specifically have really dropped the ball on this.
I interact regularly with AWS to support our needs in MLOps and to some extent GenAI. 3 of the experts we talked to have all left for competitors in the last year.
re:Invent London this year presented nothing new of note on the GenAI front. The year before was full of promise on Bedrock.
Outside of AWS, I still can’t fathom how they haven’t integrated an AI assistant into Alexa yet either
> Alexa+ costs $19.99 per month, but all Amazon Prime members will get it for free.
I'm curious if non prime members make up a big market for Alexa. I rarely use my smart devices for anything beyond lights, music, and occasional Q&A, and certainly can't see myself paying 20$/month for it.
It works out for them because bandwidth gets cheaper over time but inflation eats away at that. $70 today is like $50 back in 2010 when GFiber first launched.
Alexa consistently fails with the simplest of questions.
Only thing it can do is set a timer, turn off a light and play music.
It is still nice, but it’s so frustrating when a question pops into my mind, and I accidentally ask Alexa just to get reminded yet again how useless it is for everything but the most basic tasks.
And no, I won’t pay 240 dollars a year so that I can get a proper response to my random questions that I realistically have only about once a week.
> Only thing it can do is set a timer, turn off a light
And it can't even do that without an Internet connection. As someone who experiences annoyingly frequent outages, it never ceases to boggle my mind that I have a $200 computer, with an 8" monitor and everything, that can't even understand "set a timer for 10 minutes" on its own.
Fair question, playing music and timers are nice, but their AI is abysmal so you can imo not ask it to do anything else. I previously worked for a smart home company so I wanted to test out product’s integration with Alexa, so we have some at home. I’m planning to get rid of them, though, and only leave one in the kitchen.
Originally bought it for an elderly parent in assisted living; wasn't as useful as we'd hoped; repurposed it as a kitchen-timer/music-box, for which it works adequately as long as the Internet connection is up. I would not recommend anyone else buy one.
Being able to just order something with zero shipping has a ton of value. I could drive down the street but it would still be an hour at the end of the day.
Video streaming has some value but there are a lot of options.
I'm predicting that Grok fails simply due to half (?) the software engineering populating not wanting to use anything Musk has developed.
Grok has to be more than n-times (2x?) as good as anything else on the market to attain any sort of lead. Falling short of that, people will simply choose alternatives out of brand preference.
This might be the first case of a company having difficulty selling its product, even if it's a superior product, due to its leader being disliked. I'm not aware of any other instances of this.
Maybe if Musk switches to selling B2B and to the US government...
If you piss off half of your possible user base, adoption becomes incredibly difficult. This is why tech and business leaders should stay out of politics.
> I'm predicting that Grok fails simply due to half (?) the software engineering populating not wanting to use anything Musk has developed.
I think that's a wildly optimistic figure on your part.
Lets assume that developers are split almost 50/50 on politics.
Of that 50% that follows the politics you approve off, lets err on the side of your argument and assume that 50% of those actually care enough to change their purchases because of it.
Of the 25% we have left, lets once again err on the side of your argument and assume 50% care enough about the politics to disregard any technology superiority in favour of sticking to their political leanings.
Of the 12.5% left, how many do you think are going to say "well, let me get beaten by my competitors because I am taking a stand!", especially when the "beaten" means a comparative drop in income?
After all, after nazi-salute, mecha-hitler, etc blew up, by just how much did the demand for Teslas fall?
The fraction of the population that cares enough about these (on both sides) things are, thankfully, single-digit percentages. Maybe even less.
>>After all, after nazi-salute, mecha-hitler, etc blew up, by just how much did the demand for Teslas fall?
I had been saving up for a Tesla but now I am looking elsewhere. I think a lot of people are doing the same here in Canada. You can grok the actual numbers if you want.
Yeah, a simple example is to just look at how many companies/universities have ChatGPT vs Grok subscriptions internally. I can imagine that many people would have a problem with subscribing to Grok, even if its performance is comparable.
> This is why tech and business leaders should stay out of politics.
Yeah but they don't stay out of politics, do they? Gemini painting black Nazis was a deliberate choice to troll the vast majority of the population who isn't woke extremists.
My family subscribes to Grok and it's because of politics, not in spite of it. The answer gap isn't large today but I support Musk's goal of building a truth seeking AI, and he is right about a lot of things in politics too. Grok might well fail financially, the current AI market is too competitive and the world probably doesn't need so many LLM companies. But it's good someone wants AI to say what's true and not merely what's popular in its training set.
I think their point was that becoming very involved in politics in a way that alienates half of the population has tarnished Musk’s brand (although, I’d personally adjust that down to more like 1/3). If the point of your whataboutism is that previously it alienated the other 1/3… that doesn’t seem to improve their odds, right?
If anything they’ve now pissed off 2/3 of the population at some point or another.
Literally not. Elon Musk even published the infamous algorithm which he had claimed silenced and censored right wing voices. The only thing the algorithm did that was odd was that it had a special case written into it to boost Elon Musk. You can go look it up.
Mechahitler, the South African genocide debacle, explicitly checking Elons Twitter feed, “You get your news from infowars” system prompts, etc have basically made Grok not a real option for me. I do not want to use a product that is specifically being engineered to be a right wing disinformation machine.
And no, generic brand safety mishaps are not the same; everyone is not doing this.
I enabled Alexa+ few days ago on my devices. Everyone in our home immediately disliked the new Alexa. There were some fairly basic things that Alexa+ cannot do, and Alexa was able to do. Some fairly simple question/answering tasks, and questoins about status of an order.
They basically have with Alexa+. It's slightly more limited than ChatGPT, but it sounds much more realistic than stock Alexa and blows it out of the water in terms of smarts. The old model was basically a Siri-like "set timers and check the weather with specific commands," plus some hit-or-miss skills you had to install separately. But the new one gives much more of a sense of understanding your question and can carry on conversations with contextual responses. I've been pretty impressed with it, and the nature of the Echo device makes it much easier to query at will than having to open the ChatGPT app and switch to voice mode.
I agree. I think the Echo devices are good for certain kinds of voice-driven LLM experience. Although it's not that useful for detailed responses and serious questions, since you can't go back and read its response again.
Having briefly interacted with AWS Q out of curiosity, I can see why they haven’t pushed much out publicly. Aside from giving someone a chuckle when they decided to call its suggestions “Q Tips”, it’s functionally useless.
They all but abandoned Astro, their home robot. My suspicion (and information I've heard internally) all but points at them only using Astro as a testbed for self-navigating warehouse robotics, and now that they got what they wanted out of it, the Vesta team basically got thrown to the wolves.
Lingo question: is MLOps like devops for ML, or like flops for ML? I wonder because… actually, either case seems like somewhere Amazon might be losing experts to hot startups.
As the other response said. It’s DevOps for ML. They have Amazon SageMaker which is the managed ML/MLOps offering that we use extensively because we’re a small team. The documentation is awful
They thought Alexa will enable users to buy more from Amazon just by voice. But most users turned out like me. I would not spend a single dollar on Amazon without actually seeing the item on my mobile or desktop. I wouldn’t even add to cart via Alexa. That’s not an ideal user for device and service that requires hundreds of millions to run.
You saw this with Amazon Dash buttons too. This idea that users would just go "Hey order me some more Tide" and Amazon would just do the right thing at the right price like some sort of intelligent personal assistant. Which it by no means is.
That's because there is no lock-in in the current ecosystems for AI. Yet. But once AIs become your lifetime companion that know everything there is to know about you and the lock-in is maximized (imagine leaving your AI provider will be something like a divorce with you losing half your memory) these parties will flock to it.
The blessing right now is the limit to contextual memory. Once those limits fall away and all of your previous conversations are made part of the context I suspect the game will change considerably, as will the players.
There's a chance this memory problem is not going to be that easy to solve. It's true context lengths have gotten much longer, but all context is not created equal.
There's like a significant loss of model sharpness as context goes over 100K. Sometimes earlier, sometimes later. Even using context windows to their maximum extent today, the models are not always especially nuanced over the long ctx. I compact after 100K tokens.
But you don't have to hold the entire memory in context. You just need to perfect techniques to pull in parts of the context that you need. This can be done via RAG, multi-agent architectures, etc. It's not perfect but it will get better over time.
From my experience context window by itself tells half the story. You load a big document that’s 200k tokens and ask it a question, it will answer just fine. You start a conversation that soon enough balloons past 100k then it starts losing coherence pretty quickly. So I guess batch size plays a more significant role.
I'm over simplifying here but graph database and knowledge graphs exist. An LLM doesn't need to preserve everything in context, just what it needs for that conversation.
Unless there is a trick that I am missing, I don't think this will work by itself. The fundamental thing is what can the model attend to as it generates the next token.
If you give a summary+graph to the model, it can still only attend to the summary for token 1. If it's going to call a tool for a deeper memory, it still only gets the summary when it makes the decision on what to call.
You get the same problem when asking the model to make changes in even medium-sized code bases. It starts from scratch each time, takes forever to read a bunch of files, and sometimes it reads the right stuff, other times it doesn't.
Context will need to go in layers. Like when you tell someone what you do for a living, your first version will be very broad. But when they ask the right questions, you can dive into details pretty quick.
Export your old chats and put them in a RAG system accessible on the new LLM provider. I did it. I made my chat history into a MCP tool I can use with Claude Desktop or Cursor.
Ever since I started taking care of my LLM logs and memory, I had no issue switching model providers.
> But once AIs become your lifetime companion that know everything there is to know about you and the lock-in is maximized
Why? It's just a bunch of text. They are forced by law to allow you to export your data - so you just take your life's "novel" and copy paste it into their competition's robot.
It's never quite that straightforward, or perceived as that straightforward. That's why most people just renew their insurance as it's easier than messing about changing and worrying if it will be any better. And how easy is it to transfer emails to another provider?
Who even wants all your previous conversations taken into account for everything you do? How do you grow from never forgetting anything, making mistakes, etc?
This is highly dystopian and I sure hope this will forever just be a fantasy.
I have made 100MB of my own chat logs into a RAG memory and was surprised I didn't like using it much. Why? it floods the LLM with so much prior thinking that it loses the creative spark. I now realize the sweet spot is in the middle - don't recall everything, strategic disclosure to get the max out of AI. LLM memory should be like a sexy dress - not too long, not too short. You get the most creative outputs when you hide part of your prior thinking and let the mode infer it back.
I am not an AI enthusiast but I get what you're saying. I occasionally use ChatGPT due to Google being enshittified pretty much. I often do not like the things it tells me and I for sure do not like it complimenting everything I do, but thats something other people seem to like...
In my experience starting a fresh chat after a while of back and forth can really help, so I agree with you. Having little to zero prior context is actually the point of view one needs sometimes.
in order for that lifetime companion, we'll need to make a leap in agentic memory.
Well, let’s take your life. Your life is about 3 billion seconds (100 year life). That’s just 3 billion next-tokens. The thing you do on second N is just, as a whole, a next token. If next-token prediction can be scaled up such that we redefine a token from a part of language to an entire discrete event or action, then it won’t be hard for the model to just know what you will think and do … next. Memory in that case is just the next possible recall of a specific memory, or next possible action, and so on. It doesn’t actually need all the memory information, it just needs to know that that you will seek a specific memory next.
Why would it need your entire database of memories if it already knows you will be looking for one exact memory next? The only thing that could explode the computational cost of this is if dynamic inputs fuck with your next token prediction. For example, you must now absolutely think about a Pink Elephant. But even that is constrained in our material world (still bounded physically, as the world can’t transfer that much information through your senses physically).
A human life up to this exact moment is just a series of tokens, believe it or not. We know it for a fact because we’re bounded by time. The thing you just thought was an entire world snapshot that’s no longer here, just like an LLM output. We have not yet trained a model on human lives yet, just knowledge.
> Once those limits fall away and all of your previous conversations are made part of the context I suspect the game will change considerably, as will the players.
I dunno if this is possible; sounds like an informally specified ad-hoc statement of the halting problem.
Yeah Amazon is massively struggling to hire due to the extremely bad reputation of Andy Jasshole and the RTO 5 policy, and this is not exclusive to AI talents, but is the case for every single role. We have had reqs open for a year in my team and nobody wants to join.
Truthfully, I don't think anyone would recommend their acquaintances to join Amazon right now.
That said, Amazon is actually winning the AI war. They're selling shovels (Bedrock) in the gold rush.
I have had multiple recruiter reachouts from AWS who obviously read my resume and are interested in short cutting me into a senior role at AWS doing interesting things, but at this point AWS reputation is so bad I don’t even entertain such offers.
For senior in-demand talent you are not desperate, and really only desperate people go to work for AWS as they don’t have any better options at a company which respects their employees.
It's not like it had a good reputation earlier either (as a company, perhaps less problematic as an employer). But if I was offered multiple FAANG positions because I had some really attractive skill set, then I'd want a _lot_ more to work at Meta or Amazon than Netflix or Google, just based on my view of the corporate evilness. It's probably completely unfounded, but the fact I have that feeling just shows they haven't taken care of their brand.
Amazon having trouble to hire I think it is a well deserved result. I hope they never hire great talent again. Lately I heard they're looking for contract hires, which seems to fit their cheapness and lack of ability to attract talent.
At some point, the turnover has to lead to "the blind leading the blind" with nobody having a clearer big picture view on the software they own. This can't be a productive way to run a company, but they seem to persist nonetheless. It may take many years, but I imagine their software will rot from within due to their hiring practices.
Exactly, Amazon practices the equivalent of a decimation of their workforce. This may even work in the initial years, but over time they'll quickly lose their best minds and the software will be unmaintainable.
It's almost funny how they just don't give a shit about being an attractive employer. They never have. Going back to 2002, it's always been "if you don't like it, there's the door."
It seems that they just don't care about the high turnover.
Bedrock? It's like a vibe-coded "router" app. It really doesn't provide anything that is not provided by countless other companies.
AWS is falling behind even in their most traditional area: renting compute capacity.
For example, I can't easily run models that need GPUs without launching classic EC2 instances. Fargate or Lambda _still_ don't support GPUs. Sagemaker Serverless exists but has some weird limits (like 10GB limit on Docker images).
Bedrock is not at all a router. They do provide a routing capability now, but at its core it's a wrapper around models so you can interact with any model with the same unique API.
> For example, I can't easily run models that need GPUs without launching classic EC2 instances.
Yeah okay, but you can run most entreprise-level models via Bedrock.
Only if you want them to go to random inference regions. God forbid you would want inference in a single region. Then you need to be satisfied with 12 month old models that have been superseded 2 times already.
AWS doesn't need to do anything innovative and the enterprises still come. Every product AWS sells has a similar offering from a competitor. But businesses stick with amazon because its all in one. They get bills from one company, trust their security with one company, ect. The only thing that matters to AWS is its reputation.
This works up to a point. I'm extremely familiar with AWS, but we simply _can't_ use it to train our models because it costs 2-3 times more than their competitors. All while requiring us to basically bring up all the infrastructure around maintaining the training cluster ourselves.
Frankly, this is strictly a positive signal to me.
Fargate and lambda are fundamentally very different from EC2/nitro under the hood, with a very different risk profile in terms of security. The reason you can't run GPU workloads on top of fargate and lambda is because exposing physical 3rd-party hardware to untrusted customer code dramatically increases the startup and shutdown costs (ie: validating that the hardware is still functional, healthy, and hasn't been tampered with in any way). That means scrubbing takes a long time and you can't handle capacity surges as easily as you can with paravirtualized traditional compute workloads.
There are a lot of business-minded non-technical people running AWS, some of which are sure to be loudly complaining about this horrible loss of revenue... which simply lets you know that when push comes to shove, the right voices are still winning inside AWS (eg: the voices that put security above everything else, where it belongs).
> Frankly, this is strictly a positive signal to me.
How?
> The reason you can't run GPU workloads on top of fargate and lambda is because exposing physical 3rd-party hardware to untrusted customer code dramatically increases the startup and shutdown costs
This is BS. Both NVidia and AMD offer virtualization extensions. And even without that, they can simply power-cycle the GPUs after switching tenants.
Moreover, Fargate is used for long-running tasks, and it definitely can run on a regular Nitro stack. They absolutely can provide GPUs for them, but it likely requires a lot of internal work across teams to make it happen. So it doesn't happen.
I worked at AWS, in a team responsible for EC2 instance launching. So I know how it all works internally :)
You'd have to build totally separate datacenters with totally different hardware than what they have today. You're not thinking about the complexity introduced by the use of pcie switches. For starters, you don't have enough bandwidth to saturate all gpus concurrently, they're sharing pcie root complex bandwidth, which is a non-starter if you want to define any kind of reasonable SLA. You can't really enforce limits, either. Even if you're able to tolerate that and sell customers on it, the security side is worse. All customer GPU transactions would be traversing a shared switch fabric, which means noisy bursty neighbors, timing side-channels, etc., etc., etc.
Not having GPUs on Fargate/Lambda is, at this point, just a sign of corporate impotence. They can't marshal internal teams to work together, so all they can do is a wrapper/router for AI models that a student can vibe-code in a month.
We're doing AI models for aerial imagery analysis, so we need to train and host very custom code. Right now, we have to use third-parties for that because AWS is way more expensive than the competition (e.g. https://lambda.ai/pricing ), _and_ it's harder to use. And yes, we spoke with the sales reps about private pricing offers.
none of this applies to g6e because it doesn’t have/need a pcie switch, because it doesn’t have rdma support (nor nvlink), which means sriov just works.
Nothing, but IMO it’s a bad idea. 1. customers who build a compute workload on top of fargate have no future, newer hardware probably won’t ever support it. 2. It’s already ancient hardware from 3 years ago. 3. AWS now has to take responsibility for building an AMI with the latest driver, because the driver must always be newer than whatever toolkit is used inside the container. 4. AWS needs to monitor those instances and write wrappers for things like dgcm.
Fargate is simply a userspace application to manage containers with some ties-in to the AWS control plane for orchestration. It allows users to simply request compute capability from EKS/ECS without caring about autoscaling groups, launch templates, and all the other overhead.
"AWS Lambda for model running" would be another nice service.
The things that competitors already provide.
And this is not a weird nonsense requirement. It's something that a lot of serious AI companies now need. And the AWS is totally dropping the ball.
> AWS now has to take responsibility for building an AMI with the latest driver, because the driver must always be newer than whatever toolkit is used inside the container.
They already do that for Bedrock, Sagemaker, and other AI apps.
I do want to point out that I wrote this in 2019. It was a vastly different landscape, and AWS did a terrible job of promoting the ML value proposition.
I would love to see a retrospective on the AWS uncomfortable truths as a blog post and how they've held out over the past couple years.
2019 was a different time - though I suspect that your statement about making money (as in profit) rather than just revenue (reselling compute for less than you bought it) would hold true for most companies.
> Last year, OpenAI expected about $5 billion in losses on $3.7 billion in revenue. OpenAI’s annual recurring revenue is now on track to pass $20 billion this year, but the company is still losing money.
> “As long as we’re on this very distinct curve of the model getting better and better, I think the rational thing to do is to just be willing to run the loss for quite a while,” Altman told CNBC’s “Squawk Box” in an interview Friday following the release of GPT-5.
Selling compute for less than it cost you will have as much revenue as you want to pay for.
Anthropic founder described it as: if each model were a company, they be hugely profitable. It looks bad since when the model you trained in 2024 is generating net positive revenue, you’re also training a more expensive model for 2025 that won’t generate revenue until then. So currently, they’re always burning more cash than they’re bringing in, under the expectation that every model will increase revenue even more. Who knows how long that lasts, but it’s working so far.
Paraphrase is from the podcast he was in with the stripe founder, cheeky pints I think
The problem for OpenAI and the difference with other FAANGs is that they don’t own the internet. Other companies are able to replicate their product, which prevents them from fully realizing profits.
Google doesn’t have this problem. They only run Google ads in their search results. Same thing for Facebook.
If I have the numbers right, OpenAI will burn more money this year alone than all of those prior companies did in their entire profitless phase of existence.
AWS has always ridden other products (Postgres, MS-SQL, Redis, etc) that are open source or has negotiated licenses (Windows, MS-SQL, Oracle RDBMS) that are bundled in the end-user price per hour/GB/whatever.
AWS has Bedrock to use various AI providers and has bundled the licensing into the price, so they are getting the users without having to develop the actual AI.
They provide the compute, networking etc, and they provide the users to the AI vendors.
I do not see it as a master strategy. It is just something that happened. Similar to having no plan of having million fakeish Chinese brand selling crap on Amazon.com. But ones they are there then Sure, why not . May be in few year a lot of crashed and burned AI talent will be looking for boring corporate IT/AI job and Amazon will be around to offer that. And if does not happen there will still be ton of other work to do for Amazon.
I don't that's true. Amazon sells infrastructure to other AI companies. If they jump into the model race they become a competitor not an infrastructure provider.
I don't know who needs to hear this but, you can be a big tech company and not compete for every single market the other big tech companies are going for.
How do you define “missed”? I ask because AI is a bit different than those earlier trends: yes, investors wasted money during the dotcom bubble but there were thousands and thousands of places which spent a modest amount of money and saw almost immediate benefits. AI has had some modest wins but there hasn’t been that “customers love this and are switching to OtherCorp for it” moment - it seems like there’s probably a majority of cases where the executives hyped an AI feature and people yawned and ignored it because it was for the product owner’s CV rather than the users. (Repeat a generation later for the me-too mobile apps which started from an executive’s mandate to “be in the App Store”)
In all of these cases, the problem was losing track of what actually benefits users. AI has that problem really bad now because the infrastructure is expensive and the executive class has been sold on the idea that mass layoffs are just around the corner, and they’re pushing hard to ship before the benefits are there.
> You have to be joking about them missing the internet?
No, I'm not. Bill gates famously missed it (and/or severely underestimated the need for internet on Windows PCs) in 1994/5.
Microsoft completely missed the internet, and had to play catchup throughout 1995-1998.
> They became nearly irrelevant because of mobile and had to claw their way back. That is not faring well.
That never happened. They were in no danger at any time. The historic stock price charts, if you care to look them up, would show that the mobile threat you think there was did not even put a blip on their stock price and/or their revenue.
I see what you're saying, but I don't agree with this characterization.
(1) Internet: Netscape came out in 1994, and the internet tidal wave memo was 1995 and internet explorer came out the same year. Windows was rewritten with a focus on the networking stack, with Windows NT coming out in 1993 before the web boom. The internet's value is based on network effects and while you are right that they weren't first to market, they embraced it quickly and if they hadn't it likely would have been disastrous.
(2) Stock price: if you bought MSFT in October the year the iphone came out in 2007, you would take 6 years to break even. If you bought at the top in 2000 you wouldn't break even until 2016. This is a company that was limping along. During the mobile phone boom you'd have been better off putting your money in treasuries than in MSFT.
Yes they survived and were able to do well later. But my original point still stands: if you were running MSFT and wanted to be successful you would have embraced the internet and mobile. Deliberately sitting out a major technological innovation is not a recipe for success because the risk of ruin is very high. And the risk of becoming IBM is even higher.
>(2) Stock price: if you bought MSFT in October the year the iphone came out in 2007, you would take 6 years to break even. If you bought at the top in 2000 you wouldn't break even until 2016. This is a company that was limping along. During the mobile phone boom you'd have been better off putting your money in treasuries than in MSFT.
Using equity returns to claim a business is limping along is bizarre. They were earning $10B profit per year in the early 2000s with 20%+ profit margins, something most businesses can only dream of doing, even today.
The back-loaded vesting schedule is such blatantly cynical bullshit. It shows that they're planning to overwork you, push you to wash out, and undercompensate you for the experience, which is exactly what I've seen happen to a good number of friends. Amazon has become notorious here in Seattle - everyone knows they're a burnout factory. Some people make it through, and they make good money, but you have to really care about money for that to be worth the effort.
I had an Amazon interview loop on the calendar during my recent job search, a couple of months back, but it was difficult to get excited; they think so very highly of themselves, for what they're offering - and I don't just mean the money, but the culture too. They treat you like an interchangeable wage slave, not like a respected professional; it's all hoops to jump through, and procedures to memorize - dance, monkey, dance!
The recruiter was shocked when I cancelled the rest of the interviews, like, aren't you even going to give us a chance? But no: I had received a good offer from an ambitious, well-organized, well-funded AI startup which was excited to have me on board. With that on the table, why would I put up with Amazon? They won't offer better pay, they can't offer a better culture, and they don't have more interesting problems to work on.
This is a serious challenge in relation to hiring also. If you want to pay for good talent, and so are prepared to pay good money, how do you avoid people who are there for the money.
They got away with this attitude in the earlier days but it’s really hurting them now. A good chunk of the best talent out there won’t even consider Amazon. Culturally it’s very hard to turn that around now and catch up.
90% of the folks there that I know that were good have left for elsewhere. Of the ones that didn’t most are on H1Bs and basically have no choice but to stay and deal with the toxic environment.
This is an uninformed take. Yes the RSU is backloaded. But during the first two years, you get a large monthly cash sign up bonuses so that assuming the stock stays flat, over the four years, your total comp stays flat. If the stock increases your comp goes up.
I spoke to someone who is there now and when you get your yearly review, now you can choose between mostly cash vs mostly stock for your raise and most people choose mostly cash.
I make the same now as I did when I was at AWS and I much prefer my all cash comp over my less cash + RSUs when I was there.
RSU grants assume a growth rate (15%? I forget) so if they stay flat, go down, or grow slower than the baked-in growth rate, then you make less each year. If you do well enough, they’ll give you some RSUs to “make you whole” (as they used to say) but that doesn’t really happen anymore (or not much).
This is not true for your initial four year grant. I’m going to make up a number to make the math easy. Say my total compensation target was $200K. My initial 4 year offer was structured based on the then current stock price.
It would have been what ever it takes where base + prorated signing bonus + RSUs would equal $200K taking into account the 5/15/40/40 RSU schedule.
> The back-loaded vesting schedule is such blatantly cynical bullshit;
I don’t understand the complains about it. Amazon pays monthly cash ”sign-on bonus” in the first two years, which is ~ equal to the stock that you get in the years three and four (counting at the grant price). Is this fact not advertized well enough?
The "sign-on bonus" comes with serious strings attached. A good friend of mine got royally screwed when he mistook that bonus for real money, then got pushed to the point of burnout and had to leave; Amazon demanded a lot of the money back, but he didn't have it anymore.
I worked at Amazon in 2021 and rage-quit after 9 months. The sign-on bonus I received was paid out monthly, so I didn't have to pay anything back. If it's large enough, they pay it monthly because they know it's very likely you won't make it to the 2nd year.
Well for me, I was already 46 when a recruiter from Amazon Retail reached out to me about an SDE (software development) position at Amazon Retail. They said it would require relocation after COVID (this was April 2020). I knew about Amazon’s reputation from both stories and my best friend who had worked as an L6 in the finance department.
There was no way in hell I was going to sell my house and uproot my life to work for Amazon. Then the recruiter after she kept talking suggests I interview for a “permanently remote” [1] “field by design” role at AWS ProServe. I thought sure why not?
The plan was always to make some money - I made over a quarter million more over 3.5 years than I could have made as an enterprise dev working in Atlanta - put AWS on my resume, gain some industry contacts and move on in four years.
I saw the writing on the wall shortly before my 3 year anniversary. I played the game well enough to get past my next vesting period and get my “bust your ass and try to work through your PIP or receive a $40K+ severance and ‘leave immediately’”.
I didn’t hesitate. I took the severance and already had two job offers lined up and had been waiting on the severance offer.
[1] They forced their “field by design” customer facing roles in the office at the end of last year. I would have left anyway before I ever went back into the office.
My friend is a native-born Seattleite, so no, it was definitely not a relocation bonus.
Perhaps they recently changed their policies? I don't know, but it's not a risk I would want to take. Who would want to work for people who treated their coworkers like that?
Alright, I did some quick research and it seems that they do sometimes pay full first year of sign-on bonus, which you need to repay (prorated). I didn’t see that that during my time at Amazon.
The full payment that requires pro-rates is even worse. They expect you to pay it fully back. (ie. with the deducted taxes included!)
I bet it is possible to profit from a such scheme if Amazon is able to declare that as a reversed-transaction (similar to VAT-refunds) at the end of the fiscal year.
I worked there in 2013 and had the signing bonus paid monthly. I thought it was great since I could work there as long as I could tolerate it (10 months) and leave without regrets about having to pay back anything. Decent cash comp so I feel I got a good deal.
I joined in 2022 from a different location, there were 2 kinds of comp in terms of bonues, each split into 2 other;
1. Relocation package
a. Lump-sum (7k EUR): You get certain amount of money, and you deal with your own move yourself. (Albeit with some reimbursement possible for the initial trips)
b. "Other" (I don't remember the name): More supportive option, good if you have family & furniture to move. They essentially pay everything for you.
c. Important: The 7k EUR was subject to the tax, hence I got taxed at 55% (EU) due to having no tax residency at the moment (obviously). Nobody ever mentions this. But the re-payment is with the tax-included, ie. you are expected to pay 7k back!
2. Sign-on bonus: This splits into 2-year period
a. 1st year: 50% of the total bonus, transferred to your bank account on your first work day.
b. 2nd year: Each month, you get 1/12 of the remaining 50%, essentially something like ~4.18% each month on the second year.
c. The 50%/50% ratio may depend on the team/role/location, I heard some of the L4s joined to the team got split of 40%/60% (ie less in the first year) for reasons unbeknownst to me.
Conditions are pretty simple, if you leave (for any reason), you must repay monthly-pro-rated amount that you haven't worked given the total period is 24-months. ie. In Luxembourg, probation is 6-months. (Until) at the end of the probation, Amazon can just fire you for no reason. In this case, since the 2nd year sign-on hasn't vested yet, nothing to pay from that, but you must pay 1/4th of your "relocation expenses" and full half of (ie untaxed full amount divided by 2) sign-on bonus you receive on your first day. (ie. 25% of the total sign-on bonus)
Firstly, I know someone (a Greek national) who left Amazon during his 12th Month. Amazon demanded total of 4k+ euros from the guy, citing he hasn't finished his 12th month, hence the first half of his relocation bonus plus the 1-month of pro-rated sign-on bonus, before tax. As far as I know, it was more or less equivalent to his monthly gross salary, and he paid in installments.
Secondly, I heard someone joined from non-EU country in 2023, and essentially got laid off. But because she was in probation and obviously worker rights are much stricter in EU, Amazon just declared her as a probation-failed case instead of layoff. (She also got laid off within last 2 weeks of her 6-months long probation). Since she only got the residence permit recently, not having more than a few months (when unemployed as a 3rd-country national), plus Amazon demanded money to be paid back. As far as I know she contacted an labour lawyer and they basically advised her to go back and not to pay anything back as it becomes an international matter. And the costs/fees for such is much higher than what would Amazon get it back, hence she did what was suggested. Although it obviously burns the bridges but in this case, Amazon started the fire first...
---
As a result, the practices applied here falls no short of what you can hear from the news. As the company has no heart or soul, people are just numbers in a balance-sheet...
> The back-loaded vesting schedule is such blatantly cynical bullshit.
I don't understand this. A friend was recently offered an insane pay package from Amazon (compared to another big-tech). The way I saw it, the Amazon pay package was more attractive than the alternative because of the back-loaded vesting schedule.
Basically they pay you out in cash for the first two years, then after that you have an option to keep working there. If the stock price goes down in the first two years, you got your guaranteed cash -- no risk (and it would be a good time to interview again). If the stock price goes up, you now have basically an option on extra exposure in the form of staying longer with highly valued RSUs, and now getting some high proportion of your pay in RSUs.
It just seems straight up better? If you want the stock instead of fungible cash, just buy it on the open market?
It's bullshit because it assumes 15% IRR. So if they tell you you're getting $100K in outyear 3, it's not actually $100K, it's $65K of present value equities. If it fails to reach the target value, well, "Ownership" is an LP. You might get some more stock that vests in another year to make up for it, but that assumes you survive the PIP factory for another 12 months.
Oh, and if the stock actually goes up more than 15%, then regardless of your performance you won't get a raise because you've already exceeded band penetration.
AWS is more focused on making money off the infrastructure than on the application itself. It took same approach with kubernetes and might I say it has been very successful.
If you are a top AI researcher, there is no good reason to go to Amazon. For what? Pay? Career development? Company prospect? Work-life balance? You get nothing compared to what other companies offer.
And I say, good. We need new, smaller companies with different cultures in this space. We don't want these giant corporations to dominate and control everything.
OpenAI and Anthropic are practically subsidiaries of Microsoft and Amazon. Neither would exist without billions in cloud compute credits from their corporate benefactors. Competing in Generative AI requires the kind of resources that are only available to extremely large and established companies. I do not think all the wrapper companies count when most of them are either being bought out by the big guys or have products that are immediately outmoded in the span of months. Maybe you can make the argument for AI art companies, but Stability basically disintegrated after wasting $100m dollars and Mid journey is directly competing with Google and Meta which is not where I would want to be (aside from running a ghoulishly evil company trying to kill artistic expression).
Leaving aside the pendatic "you can't be a multiple smaller than another object", 1/5 the valuation of the 5th most valuable company in the world is probably big enough to qualify you as a big company
Market cap doesn't really feel like a good metric of anything other than what it would take to buy a company out. DuPont has a market cap of 30ish billion and 3M around 80B, and both are both larger and frankly more important than probably even Google.
Yeah, the fact that $2.5 trillion of actual investor money chose Google (Alphabet) means very little: what really matters is the opinions of anonymous commenters on HN (especially opinions that start with "doesn't really feel like")
People are so careful when writing anonymous HN comments and so careless in choosing where to invest their own money and the money of funds of which they are the professional manager
> the fact that $2.5 trillion of actual investor money chose Google
Of course, a lot of money invested in Google was invested at a much lower price; if everyone sold all at once you'd have a hard time finding 2.5T of new money to buy all those shares. We could argue about if "not selling" is the same as "choosing again at the new price" every day... but... Google's not the interesting case here anyway.
For a young company in a hot industry like OpenAI total market cap is even less relevant since so much of the company simply isn't liquid anyway and the numbers come from far fewer instances of purchases than for an established public one.
Investors look at how much money is already invested in a company in deciding whether to invest. I.e., investors pay close attention to market cap.
If Google's market cap were $25 trillion, practically nobody would buy Google stock (and practically everyone who already held the stock would immediately sell) because most investors do not believe that Google can ever pay enough dividends or buy back enough stock to justify such a high valuation.
A company's market cap is a collective estimate of how much money the company will to return to investors in the future. When the company is publicly-traded in an open informational regime such as the US, this collective estimate is usually quite "accurate" in the sense that it is very difficult for any single analyst or single team of analysts to improve on the estimate.
An investor can make a big bet on a small company, yes, but the market cap of a company is more than just an indication of how much money has been bet on the company: it also mean that every investor (big or small) who still holds the stock believes that the expected amount of money that company will return to shareholders exceeds the market cap: if there were a holder of Google stock that did not believe that, he would convert the shares into treasury bills or cash in the bank.
Amazon and Apple have never had fundamental research groups. Even before LLMs, the top big-tech fundamental labs were FAIR, Google Research and MSR.
It has never been in Amazon or Apple's DNA to chase a product that doesn't have clear revenue outcomes (as long as adoption lands). AI is no different.
IMO, it's the right decision for Amazon and wrong decision for Apple.
They got burned by over-promising Apple Intelligence and then embarking on a so-far failed, rudderless journey to land features regular people actually gave a shit about. I’m no expert, but I reckon the exact right move is concentrating on their actual deliverable products and features and letting everyone else blow their cash on a maybe months-long moat for dubiously useful advances in categories most of their customers wish everyone would stop talking about.
Beats me. My best guess is they let the hype blind them to the reality that this tech wasn’t merely a few months away from the production-level reliability they needed from it. After a while, it sank in that they couldn’t play this up as a ‘just around the corner’ release and stopped hyping up every useless beta-at-best feature like it was a huge deal. Then, when the “we’re nowhere close” internal communique was leaked, it was officially time to bow out of the hype cycle for a while.
I’m pretty sure it can’t process voice commands on-device other than whichever ones you use to activate it. The thin client architecture was the only option for most of its existence and the only one that works for older devices, which they’ve actually been great about supporting since the battery-life processor throttling bs like a decade ago. It also makes sense that they wouldn’t constantly update their architecture. Now that they’re putting their money on on-device NN stuff I imagine that will change. But let’s be real: using a smartphone without a network connection is an edge case to begin with. On-device processing makes sense for a lot of reasons, but optimizing for smartphone users without data only makes sense if you’ve got a hiking map app or something.
Apple’s biggest problem is their commitment to privacy. Delivering effective AI requires a substantial amount of user data that Apple doesn’t collect.
Their other problem is they value designers and product managers more than engineers (especially top tier AI engineers).
Both problems are basically the death knell of any hope for Apple to have good AI, but combined? It’s never gonna happen. Which is sad because Apple’s on-device hardware is quite good.
Was it an investment of actual dollars, or "just" cloud compute credits? Microsoft's investment in OpenAI was over 90% Azure credits, we're told. Which raises the possibility that the whole business is mostly about making your cloud compute business look better than it really is...
This. It's weird how most of the top tech companies are all morphing into amorphous blobs that want to get into everything and are indistinguishable from each other.
Also got to love the linguistic coincidence of Crabs and Cancer and how tech companies grow ever larger (monopolistic) to the detriment of their host (the greater economy/humanity)
My father in law was an IRS Revenue Agent. His quip was that about 20-30% of the civilian economy has tax avoidance as a primary business objective. Real estate is probably the greatest example.
Since financial engineering is in many ways more essential than the actual business. His best example was a chain hotel. In the majority of cases, a typical hotel is a tax vehicle that happens to rent rooms. So no wonder everything becomes a bank. :)
A typical chain hotel (by which I assume you mean a Marriott/Hyatt/Hilton/IHG/Choice/etc brand) is a franchised “small” business.
The franchisee typically pays 10% to 20% royalty to the franchisor (the aforementioned companies). Otherwise, they rent hotel rooms and pay staff to clean them and rent them again.
What is the tax play? That the hotel owner can 1031 into bigger and better hotels? Anyone who owns real estate can do that.
Well, one could argue the entire setup is a means of structuring investments and organizing/attracting Capital eh?
Hotel owner (aka franchisee) puts in capital in a specific way under license, gets help operating it, in exchange for the 10-20% licensing fee paid back to the main corporation.
In many cases, the owner/operator is nearly turnkey, and it’s an effective way of setting up a defacto managed business investment, almost like a LP. Many of the franchised hotels are actually owned/operated by LPs setup for the purpose.
Also in many of these cases, the franchiser provides contacts for financing, may directly facilitate/recruit Capital, and may even provide loans to the franchisee directly.
For most of these larger hotels, the actual act of renting out rooms, etc. is pretty much all automated/managed through the central system anyway, and the majority of the operating costs are structured in such a way as to minimize tax liability.
Not at all. The poster I responded to claimed this:
> a typical hotel is a tax vehicle that happens to rent rooms.
>In many cases, the owner/operator is nearly turnkey,
What does this even mean? Hotels can be turnkey, which in industry terminology means that everything is working sufficiently well such that you can start renting rooms immediately. An owner/operator being turnkey makes no sense.
> setting up a defacto managed business investment
Also makes no sense.
>Also in many of these cases, the franchiser provides contacts for financing, may directly facilitate/recruit Capital, and may even provide loans to the franchisee directly.
Even if true, what does this have to do with taxes?
>For most of these larger hotels, the actual act of renting out rooms, etc. is pretty much all automated/managed through the central system anyway,
No, the actual out of renting out rooms involves housekeepers, maintenance staff, guest service agents, cooks, and management making sure rooms are clean and habitable. Reserving a hotel room is mostly automated, but even that requires a person to manage conflicts of reservations (e.g. unexpectedly needing to extend a stay causing overbooking, changing room types, room locations, etc.)
>and the majority of the operating costs are structured in such a way as to minimize tax liability.
Who doesn't structure their operating costs to minimize their tax liability? If you file married joint instead of married separate or head of household, are you "structuring" your operating costs as a way to minimize tax liability?
The question of how a hotel is used to gain an tax advantage that would otherwise be unavailable remains unanswered.
Most properties are syndicated. Hotels are interesting because they are mix of different asset types. The GP operates the place and LPs contribute capital. Accelerated and bonus depreciation passthrough to the LPs entity.
And how is a hotel a mix of different asset types?
What does GPs and LPs have anything to do with using a hotel to gain a special tax advantage that is not available to any other commercial real estate?
I am doing research, asking the person who made the claim.
How stocks and bonds come into play is beyond me, unless I am being trolled.
But to summarize, zero evidence of how a hotel is a “tax vehicle”, nor any clarification on what a tax vehicle even is, nor why any other business wouldn’t be able to use the same strategy (if it even exists).
Dude, look up corporate partner structures. General partners. Limited partners. Etc.
Do some basic reading so you can ask informed questions from the answers you have already been given, instead of insisting someone is an idiot when they point out you are not asking useful questions.
Can you give examples? I thought becoming a bank in the US is famously difficult and regulated, so much so that most businesses who can avoid it do so by partnering up with existing, tiny banks. See almost any “fintech” solution, from startups all the way up to Apple.
As far as I understand, becoming a bank is inviting a ton of overhead with little profit potential.
I don’t think they meant a literal bank, but finance games become a bigger part of their core strategy. For example, AirBnB for a while made a majority of its profits by investing the money guests paid during the gap between booking and actual stay (paying the host).
Nit: grasses are a distinct genetic lineage, the Poaceae family. There are a few other linages outside of Poaceae that have convergently evolved to look like grasses, sedges and rushes, but they all fall in the same clade, Monocots.
Trees, on the other hand, are a growth habit, exhibited by species in a wide variety of plant families, even grasses (e.g palm trees).
There's no point in discussing a meme, but carcinisation doesn't occur in that wide of a range, and of course the reverse phenomenon (decarcinisation) is also observed.
It's a fun image, but just as Facebook isn't becoming Apple, and Amazon won't become OpenAI, evolution phenomenons are more complex than "everything becomes X"
It was common in the post wwii era in America and its Asian allies like Korea with its chaebols and Japan with its somethings I can’t remember the name of. The Asian countries forms were normally based around a single family, we’ll need more time with the current US form to see if they are also dynastic
The Japanese, family-owned, generally pre-WWII conglomerates were called zaibatsus. After WWII they were (nominally) dissolved and the now more loosely connected groups of companies are called keiretsus.
This is the unfortunate answer to a lot about why companies do
We are all addicted to growth - everyone is chasing the hockey stick curve which means a business that provides a stable business and grows modestly is seen as a failure in some parts
I don't know, there doesn't seem to be much overlap to me. Apple is a hardware business, Microsoft is software, Google is search, Facebook is social media, Amazon is distribution and compute. They do have their fingers in each other's pies but not to a large extent.
Which is a godsend for the users. Can you imagine a world where there is only one big cloud provide, say AWS, and all the big companies with the infra just sit out? Can you imagine how expensive AWS would be and how much power it has over the users?
That’s what happens when you print trillions of dollars. Suddenly investors have too much Monopoly money and they want to spend it on something, anything, that might not make as much of a loss as holding cash during the subsequent inflation.
Yeah, they will come, new companies from China, that will eat the market too, with their beautiful 996 work life balance, and we will go back to growing corn.
As a bonus you will have a very long vacation.
We, the tech, are literally a leftover of the once overwhelming engineering superiority of the west that will shrink in the next 5 years.
Just FTR - it's VERY rare for people to come up with more than one winning idea
Once a company gets big off its grand idea, there's little to no chance of it having another big winner, so buying one is best (and its cheaper too, you know it's a good idea, and you don't have to spend so much R&D on it.
Big companies have never built anything new. It goes back decades. before tech companies, it was giants like GE who grew through acquisition after acquisition and eventually imploded from the incompetence blob (which takes a long time to accumulate the damage). The same will happen to the current big tech companies in a few decades.
It makes some sense to sell out if you're building a product that will at best acquire a tiny sliver of the market, which almost all companies will. But there's at least a few AI companies, like Anthropic, that could potentially balloon towards becoming a Big Tech company. So it makes sense for them to not sell out for the time being.
You don'thave to - but that means setteling for a job that earns an okay income. Sell out for millions now - more that your lifetime earnings and use the time and money for - what you want
Most by far are working for someone else. They get no stock option, or if they get them they are of minimal value. Their generally get a good 401k (us only) and so can retire well off but would not call themselves rich.
You say this as if it's a coercive given, when you could just as easily say.. Nope, and continue to see how you compete with some agility. It might fail, but most of the big tech companies currently acquiring smaller companies themselves started small with acquisition offers being rejected along the way. Sure, there's selection bias at work there, but there are also many cases of smaller to mid-size companies that also said no to acquisition and still managed to find their successful niche.
Being acquired is not a given and neither is failure if you do compete in some way with the megacorps.
I see nothing about the current tech landscape that at all distinguishes it from previous landscapes in which smaller companies succeeded AND rejected acquisition.
I wish more people said no. However, the reality is it appears to be a given. If you're offered millions and millions of dollars, most people do not say no. The world is worse for it, but it's the truth.
I think you really need to reconsider the definition of a given. Because most people don't say no to a thing doesn't make it coercive. Whether the world is worse or better for it in this context, I can't be sure (though I lean more towards big corporations eating nearly all competitors as generally dangerous), but we're still talking about voluntary choices in a market where competition still does frequently emerge to overhaul what's established. It's impressive how often people don't notice this even as it happens all around them in ways that directly benefit their daily lives.
If that is how you feel, then the reason for why it currently is the way it is should not give you much comfort. It's not like Amazon can not decide to change things and throw more money at the issue from a different angle in the future.
But is there grounds to say that as a conglomerate they pose a large harm to market health to merit a breakup? For example, few regulators want to break up Mondragon.
There's no value in Amazon burning money to 'compete' when there no clear endgame. Right now the competition seems to be who can burn a a hundred billion dollars the fastest.
Once a use case and platform has stabilized, they'll provide it via AWS, at which poiny the SME market will eat it up.
Not only that, but all the compute spent, and hardware bought, will be worthless in 5 years.
Just the training. Training off of the internet! Filled with extremists, made up nuttery, biased bs, dogma, a large portion of the internet is stupids talking to stupids.
Just look at all the gibberish scientific papers!
If you want a hallucination prone dataset, just train on the Internet.
Over the next few years, we'll see training on encyclopedias and other data sources from pre-Internet. And we'll see it done on increasingly cheaper hardware.
This tiny branch of computer sciences is decades old, and hasn't even taken off yet. There's plenty of chance for new players.
How exactly do you foresee "pre-internet" data sources being the future of AI.
We already train on these encyclopedias, we've trained models on massive percentages of entire published book content.
None of this will be helpful either, it will be outdated and won't have modern findings, understandings. Nor will it help me diagnose a Windows Server 2019 and a DHCP issue or similar.
We're certainly not going to get accurate data via the internet, that's for sure.
Just taking a look at python. How often does the AI know it's python 2.7 vs 3? You may think all the headers say /usr/bin/python3, but they don't. And code snippets don't.
How many coders have read something, then realised it wasn't applicable to their version of the language? My point is, we need to train with certainty, not with random gibberish off the net. We need curated data, to a degree, and even SO isn't curated enough.
And of course, that's even with good data, just not categorized enough.
So one way is to create realms of trust. Some data trusted more deeply, others less so. And we need more categorization of data, and yes, that reduces model complexity and therefore some capabilities.
But we keep aiming for that complexity, without caring about where the data comes from.
And this is where I think smaller companies will come in. The big boys are focusing in brute force. We need subtle.
New languages will emerge or at least versions of existing languages till come with codenames. What about Thunder python or uber python for the next release.
(Though I'm pretty familiar with some of the concepts, I know some things to avoid (e.g., "push this button to set up a very expensive global enterprise scale observability platform of numerous complicated services, because you asked about a very simple turn-key syslog service"), and I'm expecting the occasional configuration headache (and, lately, configuration wizard bugs).)
For a new startup, I'd use AWS for all serving and hosting purposes by default, iff you have someone who can avoid pitfalls, and handle problems.
If you don't have such a technical person, maybe start off with managed Kubernetes service with high-level UI, at AWS or one of the other cloud providers, and try not to make too big a mess (which might slow you down, or take you down) before you can afford to hire specialists to make sure it keeps working for you.
I still like AWS all these years later. It’s trusted in the enterprise and you can empower people to do what they need to themselves with IAM. And it’s pretty reliable.
No one is looking at this issue correctly. Saying out of the AI "talent war" is a smart move. AI is due to collapse under its own weight.
1) High-quality training data is effectively exhausted. The next 10× scale model would need 10× more tokens than exist.
2) The Chinchilla rule. Hardware gets 2× cheaper every 18 mo, but model budgets rise 4× in that span. Every flagship LLM therefore costs 2× more than the last, while knock-off models appear years later for pennies. Benchmark gains shrink and regulation piles on. Net result: each new dollar on the next big LLM now buys far less payoff. The "wait-and-copy" option is getting cheaper every day.
I usually do not agree with the Amazon leadership (well, recently they haven't been "Right A lot"!)
But I agree with the following statement Matt Garman gave recently;
Amazon Web Services CEO Matt Garman said that using AI tools in place of junior employees was "one of the dumbest things I've ever heard" because these employees are "the least expensive" and "the most leaned into your AI tools."
It's because AI usually creates slop, without review these "slop" build up. We don't have infinite context window to solve the slop anyway. (even if we do, the context-rot has been confirmed)
Also, on average, Indian non-Tech employees who manages thousands of spreadsheets or manually manages your in-store cameras are much more cheaper than the "tokens" and the NVIDIA GPUs you can throw at the problem, at least for now and a foreseeable future.
I don't think his point was we should hire junior engineers because they're cheap and lean into AI and AI produces slop. His position is not that he wants to cheaply create slop.
He wants to hire people who are cheap and love using AI because he sees that as a better long term strategy than making senior engineers embrace AI late into their career.
from outside looking in i think this is a move i would support if i were in amazon leadership. let the other players pay for the AI movement, pick up the fruits of their labor a couple of years down the line. i dont think amazon's main play is AI anyways, if anything it's to facilitate AI with their complementary platforms in AWS
Of course not being able to monetise Alexa has always been a problem, but these and the article's issues are all to do with poor planning and top tier business direction.
Apple finance took over and scoffed at their engineers wanting to spend $10B on AI hardware, so they bought back stocks instead because Buffet understands that.
MS, Google, and Meta are spending hundreds of billions on AI, however their stocks are growing by the same amount. This doesn't seem like a crazy spending when looking at this, and better than just buying stock back and doing exactly nothing like Apple is doing.
Apple can afford to take their time, people will keep buying phones no matter whose AI they run on them. Google on the other hand can see its search business evaporating within a few years.
Of course they've stood it out. The rate of change and the R&D expenditure is off the charts. It buys them marginal utility to hire AI talent at incredible salaries to keep them at table stakes.
Meanwhile, the models are getting larger and more complex, with more users, putting the support infrastructure well beyond what individuals and even small companies can afford to outright buy. You can easily spend well over a million on even basic infrastructure to try to support some of the newer models and make it available to a few end users.
As a point of strategy for individuals and small entities, it really is cheaper in this case to spin up some AWS instances for a bit to do some LLM work and then spin them down when not in use.
So if you were AWS do you mine for gold? Or do you sell shovels?
That whole “sell shovels” thing never really made sense, even in the pre-GPU hyperscaler days. BTW, the shovel is GPU (owned by NVidia for now).
AWS, Azure, GCP weren’t just renting servers. They built whole platforms - databases, ML stacks, dev tools, security. Way more than shovels.
The moat was owning the stack. MS used Azure to power Office and now Copilot. Google used infra to juice Search, YouTube, Ads. Even Amazon used it for retail + Alexa. They were mining gold and selling shovels.
And raw compute was never where the money was. Renting VMs was the cheap layer. The profits came from all the higher level services built on top.
Now with AI it’s even more obvious:
Models drive the workloads. OpenAI/Anthropic/DeepMind aren’t just customers, they’re shaping the infra itself. Whoever owns the models sets the rules.
No models = no moat. If AWS isn’t building frontier models, it’s just reselling Nvidia GPUs while MS + Google wrap their clouds around first party models + SDKs. That pulls customers deeper into their stacks, not Amazon’s.
Falling behind compounds. Training/deploying models forces infra breakthroughs (chips, compilers, scaling). If AWS isn’t in that game, they’ll eventually struggle to even run other ppl’s models as well as rivals.
So if Amazon “sits this one out,” it’s not just losing bragging rights. It’s giving up control of the future of compute.
I agree with you that the profit comes from higher level services built on top.
But I think you are making it sound like Amazon's moat is that it came up with its own technology behind its services.
A lot of times AWS was just grabbing a bunch of popular open source stuff off the shelf and hosting it (e.g., RDS, EKS, etc). Yes there is some R&D work but almost none of what Amazon has come up with is rooted in their own work.
The value they give you is the hosting, maintenance, and compliance of all these services. If you're paying AWS extra to host your database on RDS or your Kubernetes cluster in EKS, you're generally not paying AWS to come up with a better database than anyone else, you're just paying them to help you manage permissions, backups, replication, and other maintenance/compliance/management issues that a company needs for its internal services.
In other words, Amazon's AI customers don't need Amazon to build models. They just need Amazon to use someone else's models, host them on private enterprise compute that easily ties in to existing infrastructure, RBAC, etc, and make everything compliant and easy to maintain. A whole lot of the value is being able to answer audits with "AWS handles our database backups/data security/etc" rather than saying "we have a great ops team and here's all our proof that we handle our database backups/data security/etc properly."
I think it's actually explicitly Amazon's job to sit this one out, especially since they never successfully made a good business or consumer ecosystem device like a smartphone or PC operating system.
I’m not 100% convinced this is true. Additionally, I’m not convinced that a waiting pattern right now sets Amazon up for a point of no return. It seems plausible for Amazon to pull an Apple here, to wait until technology is more mature and use their unique position to provide a quality offering.
The problem is Amazon is usually Apples non competitive cloud partner. They can’t both sit this out. AWS needs to learn in a hurry whether they should be in the model business to supply Apple with Siri LLMs. Bc if not Apple is going to Google (and Google cloud). Thats not good for AWS. Amazon is in a bit of a bind bc they should be acquiring Anthropic but not at bubble prices.
I don't understand why AWS needs to be in the model business. They didn't develop databases, they didn't develop Kubernetes, they didn't develop Linux, and the list goes on and on.
Not a whole lot in their portfolio actually has a lot of Amazon technology behind it. They've got some mild forks here and there, and they've got some stuff like Fargate that has AWS R&D work behind it but piggybacks concepts/tech stacks that definitely didn't originate from Amazon.
A lot of their value has really nothing to do with developing the underlying technology.
Is Apple really going to shovel a bunch of money to a direct competitor (Android) in a way that is likely to result in less differentiation for Apple mobile devices?
Android isn’t really a competitor it’s an ecosystem. Google Pixel is a competitor. And would Apple start funding Google’s Pixel market penetration in exchange for AI? Perhaps. Google was funding Apple with default search money for a long time. I think the 20B payment flow is about to reverse direction, not disappear.
This follows my take, in terms of where the profitability will be long term. It will be with the hardware vendors, and not the model creators. Time will see if I'm right, but, as hard as it is to create a good model, it does seem to be something that can be replicated by others.
GPUs are the land that's staked. They're of limited intrinsic value in the sense that they're in the way of getting to the gold (models), but there's a small supply of it so they go for a lot on the market. But if there was literally any other way to make the models even for a few percent cheaper, they model builders would move to that. On the inference side, most of the cloud providers are looking for pretty much any way to server that up more cheaply, with custom TPUs, or other tensor units of some type.
We saw this with crypto mining where truckloads of expensive GPUs were dumped in the trash after the proof of work became so hard it became not worth the cost of electricity to keep on that generation of card.
See, that’s the problem with what Amazon has done to you. It’s always about money with you guys. Good research is about the opposite of money. The people who don’t know what that means, who can’t fathom to understand what “the opposite of money” means without turning everything into a contrived story about money: they can’t do good R&D. Every single great R&D director will tell you this, and a bunch of people will downvote this comment, who have never been in a meaningful R&D role.
A good research culture is capable of listening to broad, generalized, completely accurate criticism in public and not downvote. Downvoting is your problem guys!
OpenAI has a million little haters out there and do you know how much time their people spend downvoting comments online? Zero. And honestly they’re paid way better than the poor souls who have wound up at Amazon, so it’s really, truly the case that none of this money money money culture really adds up to much for the little guy.
If there’s any one person to point the finger at - like why does Amazon, with its vast resources and tremendous talent, produce basically zero meaningful publicly influential research - it’s Jeff Bezos. You’re talking about strategy? The guy in charge is a colossal piece of shit, with a piece of shit girlfriend and a piece of shit world view, at least as bad as Larry Ellison, whose only redeeming factor is that MacKenzie Scott is a much smarter person than he ever was.
That’s because LLMs are just snake oil… Look at what a flop ChatGPT 5 was. If someone manages to actually make something useful, Amazon has a stake in Anthropic. Otherwise why should they waste their money while competitors bankrupt themselves at scale over a hype cycle.
> But actually, Amazon, Apple etc aren't natural homes for this, they don't need to burn money to chase it.
Why wouldn't consumer AI be a natural home for Apple?
Apple is constantly under blast for being slow to AI but if you look at the current state of AI, it feels like something Apple would never release -- the quality just isn't there. I don't necessarily think Apple only dipping their toes into AI is that poor of a decision right now. They still have the ability to blow the roof off the market with agents and device integration whenever the tech is far enough along to be trustworthy to the average consumer.
Apple's natural home is hardware, and the consume software integrated with that hardware. Off the top of my head I can't think of any "hard" software created by Apple, it's all about the UX and the integrations.
So unless Apple thinks it can outcompete it's BigTech competitors in something it historically hasn't done much of, best leave it to them.
> Off the top of my head I can't think of any "hard" software created by Apple, it's all about the UX and the integrations.
This sounds like you’re either unfamiliar with what software they make or underestimate the complexity of things like a modern operating system. For example, most people would consider Swift hard, or the various Core frameworks, or things like designing a new modern file system and doing in place migrations on billion devices, etc.
creative talents needs more comfortable and unconstrained working environment.But that against and broke many Amazon‘s leadership principal.
Also amazon dont want to pay more money to those expensive ai talents.
AI directly threatens and can enhance revenue for companies like Google and Meta, so it makes sense for them to invest in those areas. It's relevant for user segmentation, ad targeting, content creation/user engagement, and even search. LLMs are fantastically powerful search and discovery tools.
Amazon though, sells physical goods and access to physical servers. Whatever is going on with AI, Amazon will profit from without having to burn its own money in advancing SOTA.
> Amazon doesn't provide useful tools for building durable multi-AZ applications. Most customers are not going to implement Paxos
Don't really agree here, yes they screw you financially on cross-AZ bandwidth, but all of their popular services are built to work well across availability zones.
Most people don't need access to a low level consensus service like Paxos, instead they will be using one of amazons managed database services, or s3, that provides higher level abstractions, and automatically manages consensus behind the scenes.
As far as I understand it, Anthropic is effectively “Amazon AI” similar to how OpenAI is “Microsoft AI”. So they are not sitting it out, they’re very involved.
Well Amazon is mostly guided by pragmatism rather then than hype, so there they are, waiting for the dust to settle to see what directly helps their bottom line.
Amazon still has huge R&D spends, always had. Bezos had a dictum around having a high experimentation (and failure) rate as a matter of principle. They may not be making news-making moves, but I'm sure they'll develop the muscle in AI. Probably - just really honing on the customer use cases and working backwards over the long term.
It is interesting but I think they're doing the right thing. AWS bedrock works pretty well and you can access frontier models plus everything open source on it. In the end, I was disbelieving that Graviton would be good but the latest r8g series are great for compute so I imagine GPU compute will similarly be mastered by them in time.
AI will subvert and destroy Amazon's internal management culture, where status is gatekept by who can write the best 6-page reports to read before the meeting!
Or more likely -- Amazon management knows just how hard writing actually is, how hard to produce something with clarity and signal instead of just common-knowledge cliches, and so they understand that this LLM wave is overhyped. They're letting the other big players do the hard work, and effectively selling LLMs short by abstaining from the race.
I think Amazon and Apple see who is doing the "work" in commerce and manufacturing and they know and realize that some non deterministic AI is not a big threat. Sure it creates nice text, video or image but that is not "work" for these small company eating giants. They know that work counts with real goods moving in the real world, energy moving and robots that can actually act with certainity (99,999% time like internet, web as a tech ?)
Interesting theory, but Amazon uses tons of stochastic methods (including deep reinforcement learning) throughout the business, including warehouse inventory management. "Determinism" is not some north star that operations people always adhere to, because the physical world is deeply stochastic and pretending it isn't does not make for a successful operations career.
There’s Gaussian and fractal randomness. Fraud and transportation losses are Gaussian, for example - they average out to known values. An empowered LLM can wreck absolute havoc, and if it’s not empowered there’s no reason to spend $100b on training it.
This really isn’t highlighted enough. Most real world probabilities that are evaluated follow a Gaussian structure. LLMs…don’t? Fractal probably? Heavy tailed maybe (like a Cauchy distribution)? But certainly not in ways that companies are currently accustomed to.
I think Apple is just sitting idly and waiting for AI to mature to "just works" level without all the potential legal and PR minefields. It's too wild and unpredictable of an experience for their buttoned down, bland and inclusive corporate image. Apple may soon find itself producing very capable but dumb bricks if they don't catch up. Google can and will go all out on AI in Android at some point.
Amazon I think just hasn't understood how to cohesively integrate AI into their offerings. Meanwhile they're selling shovels to the prospectors with AWS.
I guess both of these understand the Ai moat is not very large, and don't buy into AGI dreams.
It has much better utility in a phone (accessibility to camera and photos, various sensors, contacts, chats, smart home, payment methods...) than on a PC.
I can imagine an AI that's more proactive, I don't go to ask a question but it helps me manage my day effectively and get more information where its useful.
Okay, but does it need to be deeply integrated into the OS or can it just interact with programs through their normal interfaces?
The most effective way to get an LLM to control a computer right now is to just give it a unix terminal because it's already a text-based environment where programs are expected to be highly interoperable.
What I'm saying is that you don't need to stop everything to redesign around AI, just allow for a decent level of interoperability that iOS (and largely android) doesn't currently have.
The mobile app development model is oriented around packaging somewhat useful software (that could usually be a web app) with malware and selling it for $0.99, necessitating a ton of sandboxing and preventing this type of interoperability in the first place. I would say focus on the semantic HTML aspect of the web and design some way for LLMs to interact with websites in an open-ended way.
A reasonable theory. Apple does hardware and supply chains, and sees how far there is to go. Nvidia does hardware too, but it's profiting hugely from the AI boom and has no reason to push back.
How do you explain the Elon keiretsu, though? Tesla and SpaceX are pretty tethered to the physical world, and in theory should have visibility into the same discrepancies that Apple sees. So why is Elon pushing so hard to develop Grok? Is it just ideology for him, or what?
Elon, like everyone, is smart at some things and dumb at others. When you realize that about the world, it will help you learn from the smart sides of folks.
Despite his mad and destructive social and political side, as an engineer and business man he is extremely smart and effective.
He makes lots of unnecessary major and cringy mistakes in both engineering and business too, but his net on both counts is astounding.
And while he may overuse it for PR, he has put himself at great financial risk when pushing through major capability developments and business hurdles. His rewards were earned.
But the sick picture of the richest person in the world, spamming stupidity, and harming countless numbers of people's lives in order to prop up his juvenile ego is hard to look past for many. For good reason.
He is a strong mix of both extremes of capability/impact spectrum, not just one.
So why did Bezos get nowhere with Blue Origin despite throwing more money at it? Or every car manufacturer that tried to build EVs before Tesla? Or every satellite internet provider before Starlink?
> Shotwell had lunch with a co-worker who had just joined the then-startup company SpaceX. They walked by the cubicle of CEO Elon Musk. “I said, ‘Oh, Elon, nice to meet you. You really need a new business developer,’” Shotwell recalls. “It just popped out. I was bad. It was very rude.” Or just bold enough to capture Musk’s attention. He called her later that day in 2002 and recruited her to be vice president of business development, his seventh employee.
Can you imagine something like that working today?
IMO Grok is the downstream consumer result of internal investment in AI at Twitter. One of the first things Elon did after buying was put all the useful APIs behind a paywall, which would be a reasonable first step if you bought it in part for the enormous training data the platform generates every day and wanted to limit competitors' access to it. Grok is then mostly just a way to get feedback on the tech.
You have all these labs spending billions on researchers and training clusters, not seeing much return on investment, meanwhile Amazon just partners with the labs, and provides inference for their models, and that seems to be fairly profitable for them.
Why would Amazon want to invest in AI, which would help their customers find what they want to buy, when they make all their profit from showing their customers what they don’t want to buy with monetized search?
The top researchers published enough details on how to build what works well. Amazon can copy what's useful. They'll probably do it in a way that makes profit, too. Neither talent wars nor AI, startup models contribute to that.
Amazon just has to host LLaMa and Qwen locally, just as they do so many other packages developed by others, and charge for their AWS compute credits. Why do they need “AI talent”?
> The company has flagged its unique pay structure, lagging AI reputation, and rigid return-to-office rules as major hurdles.
No mention of reputation for harsh/ruthless/backstabby management practices towards employees (including for tech white collar, not just biz and blue collar)?
Is that not a major factor? Or are they not aware of it? Or is mentioning it politically off-limits? Or is putting it in writing a big PR risk? Or is putting it in writing a big legal risk?
I know Amazon's reputation for treating employees poorly came up in multiple discussions at one university's big-name AI lab, for example. Not only do some people read the news, but people talk, in groups and privately.
After reading so many horror stories (whether actually true or not), my mind now just associates working at Amazon with mostly negatives: They're going to ride you like a horse and beat you up, for below-average compensation, and then if you want to claw your way up, it's a Game Of Thrones style slugfest with few winners. The opposite of "Rest and Vest." If this is exaggerated, they sure aren't doing any PR work to deny it or counter this negative reputation.
I’ve heard multiple recruiters (from different agencies in different geos) refer to them as “amholes” and said they’re hard to place and difficult to break their bad habits.
I've bumped into a lot of execs who say they don't want to hire ICs or managers (usually only one or the other) coming from specific big-name companies, and will instruct external and internal recruiters/HR and hiring managers about that.
Not big-name companies in general, but specific companies among them.
It seems to be about belief of culture taint risk (e.g., the way engineering is done, or the misaligned careerism or sharp-elbowedness that's promoted by the company). Though there's also sometimes a belief that particular large companies hire lots of people who aren't good (only, apparently, at LeetCode interviews).
I'm a bit sympathetic to those theories, though I personally don't rule out any individual. I think, say, all the FAANGs do also have individual people who are capable and well-intentioned, and haven't been permanently branded with whatever problematic culture of the company they're at.
(Though there was a time when I thought a person wouldn't have gone to one particular social media company unless they were either a sociopath or completely unaware of news in the real world, but it's more nuanced now. And there's currently an aggressively pro-fascism company that AFAICT never should've seemed like a good idea to anyone who wasn't evil or oblivious, though, I have to remember that they like to hire "impressionable children", and we now have tech track undergrads who haven't had time for anything but STEM classes and LeetCode since early teens, so they might be forgiven. I was recently considering denylisting anyone who'd gone to a different tech company, which had a well-known decades-long history of chronic underhandedness, but then I saw that a colleague who'd majorly helped me out once had finally gone there. Which is another lesson to myself not to generalize in ways unfair to the individual.)
If you want to hire people who share your values and your values include moral responsibility for the megacorp one works for, you’re right not to hire from companies you feel are immoral.
I personally don’t ascribe corporate amorality (as opposed to immorality) to all who work for it and thus with narrow exceptions would blacklist someone for working at a company who, e.g., has a CEO I dislike, practices wage suppression, etc.
I would strongly consider against hiring someone who worked in certain addiction based industries such as tobacco or gaming (not gamedev, the other kind).
Hiring ex-Amazon managers credibly signals to capital markets that a tech company past its hype phase is getting serious about controlling costs and disciplining lazy or entitled engineers. It’s in their interest to have this reputation.
> they sure aren't doing any PR work to deny it or counter this negative reputation.
They don't seem to give a shit. In the retail space their name means "low quality Chinese counterfeit products with fake reviews" and I've seen no effort on Amazon's part to counter that perception either.
As an ex-Amazonian, I hate seeing this corporate euphemism. We would be reminded yearly that compensation at Amazon was “peculiar”, when really it was just relatively low for FAANG. I would have preferred frank honesty, which I think would look like “we pay relatively low wages, for relatively good engineers, and the difference makes more money”
That used to be the case but as of a couple years ago (maybe 2023?) pay packages got bumped and in terms of TC Amazon is very competitive now. You'll likely get a better offer than Google in cash value. But the non-TC benefits are really really bad (no free food, 5 day RTO, oncall policies, etc). For those reasons I think most would take a Meta or AI lab offer over Amazon right now if they're willing to grind.
Interesting, one would think that would mean easier interviews and whatnot so as to allow for greater number of applicants and churn, but it is not what I have heard about it.
it is a good place for new graduates to solve some challenging engineering problems at scale and learn. Most of the employees do not last more than 2 years. People who stick for longer, admire that type of culture and are made for amazon. Their stock has also performed extremely well.
Under Andy Jassy’s watch Amazon really just missed the boat on AI big time. Alexa was a huge missed opportunity. They had a huge foothold and then basically did nothing with it. AWS doesn’t really have compelling AI offerings. Bedrock should be good but is a mess. The GPU offerings struggle to keep pace with competitors.
The rambling answer to the “why are you behind” question on the last earnings call indicates it’s a sore spot for leadership, but at this point it’s too little too late. The best talent has already settled elsewhere. The only real saving grace is that if/when the AI bubble pops being so far behind might not be a terrible thing.
Amazon, and AWS specially, just don’t have recognized leaders in this space at the helm. I think that’s OK as they should focus on the more boring but important infrastructure stuff.
Jassy’s long rambling answer on the last earnings call though does suggest that being way behind on AI is a sore spot for leadership.
Attracting top talent though is a challenge for Amazon beyond just AI. Their reputation has become a real issue and the top folks simply have better options.
My head hurts reading a lot of this AI-like drivel. All I see is a bunch of options for better search engines here. That said, with all this crap feeding off itself, it's not going to get any better.
I find AWS extremely difficult to use compared to GCP. Even though we received startup credits—which are essentially free money—we’re letting them go to waste because the platform is so much harder to work with.
It’s no surprise that AWS’s revenue growth is lagging behind GCP and Azure.
Beyond the AI talent gap, Amazon seems to be making serious missteps in its own core business.
It reminds me of Apple. At first, people thought Apple was being strategic by staying out of the AI race and waiting to pick the winner. But in reality, it turned out to be an inability to adapt to the new trend. I expect the same pattern from Amazon.
Zuckerberg rushing into every new fad with billions of dollars has somehow tricked people into thinking that's what big tech is about and all of them should be shovelling money into this.
But actually every other company has been much more strategic, Microsoft is bullish because they partnered up with OpenAI and it pumps their share price to be bullish, Google is the natural home of a lot of this research.
But actually, Amazon, Apple etc aren't natural homes for this, they don't need to burn money to chase it.
So there we have it, the companies that have a good strategy for this are investing heavily, the others will pick up merges and key technological partners as the market matures, and presumably Zuck will go off and burn $XB on the next fad once AI has cooled down.
I generally agree with you, although Amazon is really paranoid about being behind here.
On the last earnings call the CEO gave a long rambling defensive response to an analyst question on why they’re behind. Reports from the inside also say that leaders are in full blown panic mode, pressing teams to come up with AI offerings even though Amazon really doesn’t have any recognized AI leaders in leadership roles and the best talent in tech is increasingly leaving or steering clear of Amazon.
I agree they should just focus on what they’re good at, which is logistics and fundamental “boring” compute infrastructure things. However leadership there though is just all over the map trying to convince folks their not behind vs just focusing on strengths.
Doesn't Amazon have a huge lead just because of AWS? Every other player is scrambling for hardware/electricity while Amazon has been building out data centers for the last 20 years.
> Doesn't Amazon have a huge lead just because of AWS?
They have huge exposure because of AWS; if the way people use computing shifts, and AWS isn't well-configured for AI workloads, then AWS has a lot to lose.
> Every other player is scrambling for hardware/electricity while Amazon has been building out data centers for the last 20 years.
Microsoft and Google have also been building out data centers for quite a while, but also haven't sat out the AI talent wars the way Amazon has.
> AWS isn't well-configured for AI workloads
What does that mean? Not enough GPUs?
A few things, including:
1. Price-performance has struggled to stay competitive. There’s some supply-demand forces at play, but the top companies consistently seem to strike better deals elsewhere.
2. The way AWS is architected, especially on networking, isn’t ideal for AI. They’ve dug their heels on in their own networking protocols despite struggling to compete on performance. I personally know of several workloads that left AWS because they couldn’t compete on networking performance.
3. Struggling on the managed side. On paper a service like Bedrock should be great but in practice it’s been a hot mess. I’d love to use Anthropic via Bedrock, but it’s just much more reliable when going direct. AWS has never been great at these sort of managed services at scale and they’re again struggling here.
In theory they should, but it’s increasingly looking like they’re struggling to attract/retain the right talent to take advantage of that position. On paper they should be wiping the floor with others in this space. In practice they’re getting their *ss kicked and in a panic on what to do.
My understanding is that they fell behind on offering the latest gen Nvidia hardware (Blackwell/Blackwell Ultra) due to their focus on internally developed ASICs (Trainium/Inferentia gen 2).
Which bar raiser is going to raise the bar first??
I'd argue that Meta's income derives in no small part from their best in class ad targeting.
Being on the forefront of
(1) creating a personalized, per user data profile for ad-targeting is very much their core business. An LLM can do a very good job of synthesizing all the data they have on someone to try predicting things people will be interested in.
(2) by offering a free "ask me anything" service from meta.ai which is tied directly to their real-world human user account. They gather an even more robust user profile.
This isn't in-my-opinion simply throwing billions at a problem willy nilly. Figuring out how to apply this to their vast reams of existing customer data economically is going to directly impact their bottom line.
5 minutes on facebook being force-fed mesopotamian alien conspiracies is all you'll need to experience to fully understand just how BADLY they need some kind of intelligence for their content/advertising targeting, artificial or not...
Obviously one is a very bad sample, but why are the ads I see on FB so badly targetted?
You probably don't spend enough time on their sites to have a good ad targeting model of you developed. The closer you are to normal users, with hundreds of hours of usage and many ad clicks, the more accurate the ads will be for you.
You mean the closer I am to the top of the bell curve, the more your ads "shooting from the hip" will land? Who would've thunk it?!
Did you block their tracking across the whole damn internet, by any chance?
Same terrible experience for me while I was on FB. I was spending a lot of time there and I do shop a lot online. They couldn’t come with relevant ad targeting for me. For my wife they started to show relevant ads AFTER she went to settings and manually selected areas she is interested in. This is not an advanced technology everyone claim FB has.
Instagram has killer ad targeting; no wonder all these direct-to-consumer brands flock there. FB not so much I agree.
>An LLM can do a very good job of synthesizing all the data they have on someone to try predicting things people will be interested in.
Is synthesizing the right word here?
I think is absolutely is, LOL. Though a "very good job of synthesizing" might not actually good for much...
People look at all the chaos in their AI lab but ignore the fact that they yet again beat on earnings substantially and directly cited better ad targeting as the reason for that. Building an LLM is nice for them, but applying AI to their core business is what really matters financially, and that seems like it's going just fine.
The largest LLMs are mostly going to be running in the cloud, so the general purpose cloud providers (Amazon, Microsoft, Google) are presumably going to be in the business of serving models, but that doesn't necessarily mean they need to build the models themselves.
LLMs look to be shaping up as an interchangeable commodity as training datasets, at least for general purpose use, converge to the limits of the available data, so access to customers seems just as important, if not more, than the models themselves. It seems it just takes money to build a SOTA LLM, but the cloud providers have more of a moat, so customer access is perhaps the harder part.
Amazon do of course have a close relationship with Anthropic both for training and serving models, which seems like a natural fit given the whole picture of who's in bed with who, especially as Anthropic and Amazon are both focused on business customers.
Microsoft is building it's own in-house LLM's based on OpenAI's IP. Google builds it's own models.
Sure, but you can also sell something without having built it yourself, just as Microsoft Copilot supports OpenAI and Anthropic models.
It doesn't have to be either/or of course - a cloud provider may well support a range of models, some developed in house and some not.
Vertical integration - a cloud provider building everything they sell - isn't necessarily the most logical business model. Sometimes it makes more sense to buy from a supplier, giving up a bit of margin, than build yourself.
I'm just an observer. Microsoft has invested billions in OpenAI and can access their IP as a result. It might even be possible MS hopes that OpenAI fails and doesn't allow them to restructure to continue to acquire outside funding. You can go directly to the announcement of their in-house model offerings and they are clearly using this as a recruiting tool for talent. Whether it makes sense for the cloud providers to build their own models is not for me to say, but they may not have a choice given how quickly OpenAI/Anthropic are burning cash. If those two fail then they're essentially ceding the market to Google.
I think this analysis is a bit shallow with regard to Metas product portfolio and how AI fits in.
Much more than the others, metter runs a content business. Gen AI aides in content generation so it behooves them to research it. Even before the current explosion of chatbots, meta was putting this stuff into their VR framework. It's used for their headset tracking and speech to text is helpful for controlling a headset without a physical keyboard.
You're making it sound like they'll follow anything that walks by but I do think it's more strategic than that.
Zuckerberg bought Whatsapp and Instagram. For normal people, those replaced 90% of the internet here in Argentina
(The other 10% is mostly Google Maps and MercadoLibre.)
But that didn't require deep insight. Both were already really popular and clearly a threat to Facebook. WhatsApp was huge in Europe before they bought (possibly other places as well).
Buying competition is par for the course for near-monopolies in their niches. As long as the scale differences in value are still very large, you can avoid competition relatively cheaply, while the acquired still walk away with a lot of money.
Why does investing in AI require deep insight? ChatGPT is already huge, significantly bigger than Whatsapp when the deal was done. And while OpenAI is not for sale, he figured that their employees are. Also not to mention, investors are very positive for AI.
So far there hasn't been a transformative use case for LLMs besides the straightforward chat interface (Or some adjacent derivative). Cursor and IDE extensions are nice, but not something that generates billions in revenue.
This means there's two avenues:
1. Get a team of researchers to improve the quality of the models themselves to provide a _better_ chat interface
2. Get a lot of engineers to work LLMs into a useful product besides a chat interface.
I don't think that either of these options are going to pan out. For (1), the consumer market has been saturated. Laymen are already impressed enough by inference quality, there's little ground to be gained here besides a super AGI terminator Jarvis.
I think there's something to be had with agentic interfaces now and in the future, but they would need to have the same punching power to the public that GPT3 did when it came out to justify the billions in expenditure, which I don't think it will.
I think these companies might be able to break even if they can automate enough jobs, but... I'm not so sure.
Whatsapp had $10M revenue when it was acquired[1]. Lots of so called "chatgpt wrappers" has more revenue than that. While in hindsight Whatsapp acquisition at $19B seems no brainer, no concrete metric pointed to that compared to him investing $19B in AI now.
[1]: https://www.sec.gov/Archives/edgar/data/1326801/000132680114...
Dude Zuckerberg bought whatsapp because FB Messenger was losing market share... nothing to do with Whatsapps revenue! Rather Zuckerbergs fear of FB products being displaced.
> Cursor and IDE extensions are nice, but not something that generates billions in revenue.
I mean Cursor is already at $500 million ARR...
How many software engineers are there in the world? How many are going to stop using it when model providers start increasing token cost on their APIs?
I could see the increased productivity of using Cursor indirectly generating a lot more value per engineer, but... I wouldn't put my money on it being worth it overall, and neither should investors chasing the Nvidia returns bag.
Pretty sure everyone was balking at the purchase prices at the time
In the UK it was an obvious purchase - whatsapp was the main mode of communicaton on a phone. Nobody used Messenger for instance.
Amazon strategy is to invest in the infrastructure, money is where the machines live. I think they just realized none of those companies have a moat, so why would they? But all of them will buy compute
Except they’re struggling here. The performance of their offerings is consistently behind competitors, particularly given their ongoing networking challenges, and they’re consistently undercut on pricing.
For Amazon “renting servers” at very high margin is their cash cow. For many competitors it’s more of a side business or something they’re willing to just take far lower margin on. Amazon needs to keep the markup high. Take away the AWS cash stream and the whole of Amazon’s financials start to look ugly. That’s likely driving the current panic with its leadership.
Culturally Amazon does really well when it’s an early mover leader in a space. It really struggles, and its leadership can’t navigate, when it’s behind in a sector as is playing out here.
Under what scenario does Amazon lose the beast that is its high margin cloud service renting? It appears to be under approximately zero threat.
Companies are not going to stop needing databases and the 307 other things AWS provides, no matter how good LLMs get.
Cheaper competitors have been trying to undercut AWS since the early days of its public availability, it has not worked to stop them at all. It's their very comprehensive offering, proven track record and the momentum that has shielded AWS and will continue to indefinitely.
AWS is losing marketshare to Azure and GCP. This is big deal, it was unexpected after years of Google/Microsoft trying and failing.
Further AWS is losing share at a time when GCP and Azure are becoming profitable businesses, so no longer losing money to gain market share.
It’s already playing out. Just look at recent results. While once light years ahead competitors are now closing ranks and margins are under pressure. AWS clearly isn’t going away, but on the current trajectory its future as the leading cloud is very much not a certainty.
Because if LLM inference is going to be a bigger priority for the majority of companies, they're going to go where they can get the best performance to cost ratio. AWS is falling behind on this. So companies (especially new ones) are going to start using GCP or Azure, and if they're already there for their LLM workloads, why not run the rest of the infrastructure there?
It's similar to how AWS became the de-facto cloud provider for newer companies. They struggled to convince existing Microsoft shops to migrate to AWS, instead most of the companies just migrated to Azure. If LLMs/AI become a major factor in new companies deciding which will be their default cloud provider, they're going to pick GCP or Azure.
Except for spending cloud budgets on LLMs elsewhere like other mentioned, LLM coding will make it easier to convert codebases from being AWS dependent, easing their lock-in
I would be surprised if a cloud market leader thinks winning on commodity vm rental is a strategy
And electricity.
Microsoft has the pleasure of letting you pay for your own hosted GPT models, Mixtral, etc
Microsoft's in a sweet spot. Apple's another interesting one, you can run local LLM models on your Mac really nicely. Are they going to outcompete an Nvidia GPU? Maybe not yet, but they're fast enough as-is.
> But actually, Amazon, Apple etc aren't natural homes for this, they don't need to burn money to chase it.
Amazon is the biggest investor of AI of any company. They've already spent over $100b YTD on capex for AI infrastructure.
This is "shovels" they rent out; very different then research investment.
To do what for that money? Write summaries of product reviews? If they wanted to do something useful, they'd use the LLM to figure out which reviews are for a different product than what is currently being displayed.
"useful" means different things
> But actually, Amazon, Apple etc aren't natural homes for this, they don't need to burn money to chase it.
I really liked the concept of Apple Intelligence with everything happening all on device, both process and data with minimal reliance off device to deliver the intelligence. It’s been disappointing that it hasn’t come to fruition yet. I am still hopeful the vapor materializes soon. Personally I wouldn’t mind seeing them burning a bit more to make it happen.
It will likely occur, just maybe not this year or next. If we look over the last eighty years of computing, the trend has been smaller and more powerful computers. No reason to think this won’t occur with running inference on larger models.
Exactly. Being a tech company doesn't mean you need to do everything any more than just because you're a family doctor you also should do trauma surgery, dentistry, and botox injections. Pick a lane, be an expert in it.
Except that Amazon's AWS business is severely threatened by the rise of alternative cloud providers who offer much more AI-friendly environments. It's not an existential topic yet, but could easily turn into one.
Zuckerberg knows what he's doing, his net worth is >250 billion dollars now.
Go all in the new fad, investors pile up on your stock, dump, repeat...
> Zuckerberg knows what he's doing, his net worth is >250 billion dollars now.
Does he have this net worth because what he is doing or despite what he is doing? :-)
Correlation does not imply causation. Attribution is hard.
By that logic, a social media company shouldn't rush into it either, but they did anyway.
Zuckerbergs AI "strategy" seems to be to make it easy for people to generate AI slop and share it on FB thus keeping them active on the platform. Or to give people AI "friends" to interact with on FB, thus keeping them on the platform and looking at ads. It's horrifying but it does make business sense (IMHO) at least at first glance.
They have out sensors though for any AGI, because AGI could subvert buisness fields and expertise moats. Thats what most AI teams are- vanity projects and a few experts calming the higher ups every now and then with a "its still just autocompletion on steroids, it can not yet do work and research alone."
> Zuckerberg rushing into every new fad with billions of dollars has somehow tricked people into thinking that's what big tech is about and all of them should be shovelling money into this.
Zuckerberg failed every single fad he tried.
He's becoming more irrelevant every year and only the company's spoils from the past (earned not less by enabling, for example, a genocide to be committed in Myanmar https://www.pbs.org/newshour/world/amnesty-report-finds-face...) help carry them through to the series of disastrous idiotic decision Zuck is inflicting on them.
- VR with Oculus. It never caught on, for most people who own one, it's just gathering dust.
- Metaverse. They actually spend billions on that? https://www.youtube.com/watch?v=SAL2JZxpoGY
- LLAMA is absolute trash, a dumpster fire in the world of LLMs
Zuck is now trying to jump again on the LLM bandwagon and he's trying to...buy his way in with ridiculous pay packages: https://www.nytimes.com/2025/07/31/technology/ai-researchers.... Why is he so wrong to do that, you might ask?
He is doing it at the worst possible moment: LLMs are stagnating and even far better players than Meta like Anthropic and OpenAI can't produce anything worth writing about.
ChatGPT5 was a flop, Anthropic are struggling financially and are lowering token limits and preparing users for cranking up prices, going 180 on their promises not to use chat data for training, and Zuck, in his infinite wisdom, decides to hire top AI talent for premium price at a rapidly cooling market? You can't make up stuff like that.
It would appear that apart from being an ass kisser to Trump, Zuck shares another thing with the orange man-child running the US: a total inability to make good, or even sane deals. Fingers crossed that Meta goes bankrupt just like Trump's 6 banrkruptcies and then Zuck can focus on his MMA career.
I've been taking heat for years for making fun of the metaverse. I had hopeful digital landlords explain to me that theyll be charging rent in there! Who looked at that project and thought it was worth anything?
> I've been taking heat for years for making fun of the metaverse
I don't know in what circles you're hanging out, I don't know a single person who believed in the metaverse
> I don't know in what circles you're hanging out, I don't know a single person who believed in the metaverse
Oh please, the world was full of hype journalists wanting to sound like they get it and they are in it, whatever next trash Facebook throws their way.
The same way folks nowadays pretend like the LLMs are the next coming of Jesus, it's the same hype as the scrum crowd, the same as crypto, nfts, web3. Always ass kissers who cant think for themselves and have to jump on some bandwagon to feign competence.
Look at what the idiots at Forbes wrote: https://www.forbes.com/councils/forbestechcouncil/2023/02/27...
They are still very influential, despite having shit takes loke that.
Accenture still think the Meta is groundbreaking: https://www.accenture.com/us-en/insights/metaverse
What a bunch of losers!
71% of executives seemed to be very excited about it: https://www.weforum.org/stories/2022/04/metaverse-will-be-go...
Executives (like Zuck) are famous for being rather stupid so if they are claiming something, you bet its not gonna happen.
Apparently, "The metaverse is slowly becoming the new generation’s digital engagement platform, but it’s making changes across enterprises, too."
https://www.softserveinc.com/en-us/blog/the-promise-of-the-m...
i don't care about virtual real estate, but VR mini golf sure is fun!
meta made $62 billion dollars last year. Mark burns all this money because his one and only priority is making sure his company doesnt become an also ran. The money means nothing to him
Yet his company and him are becoming rapidly irrelevant.
So does Pichai... Every time there is something new, he forces Google to pivot, upending everything without much to show for it.
Google basically invented modern AI (the 'T' in ChatGPT stands for Transformer), then took a very broad view of how to apply broadly neural AI - AlphaGo, AlphaGenome being the kind of non-LLM stuff they've done).
A better way to look at it is that the absolute number 1 priority for google since they first created a money spiggot throguh monetising high-intent search and got the monopoly on it (outside of Amazon) has been to hold on to that. Even YT (the second biggest search engine on the internet other than google itself) is high intent search leading to advertising sales conversion.
So yes, google has adopted and killed lots of products, but for its big bets (web 2.0 / android / chrome) it's basically done everything it can to ensure it keeps it's insanely high revenue and margin search business going.
What it has to show for it is basically being the only company to have transitioned as dominent across technological eras (desktop -> web2.0 -> mobile -> (maybe llm).
As good as OpenAI is as a standalone, and as good as Claude / Claude Code is for developers, google has over 70% mobile market share with android, nearly 70% browser market share with chrome - this is a huge moat when it comes to integration.
You can also be very bullish about other possible trends. For AI - they are the only big provider which has a persistent hold on user data for training. Yes, OpenAI and Grok have a lot of their own data, but google has ALL gmail, high intent search queries, youtube videos and captions, etc.
And for AR/VR, android is a massive sleeping giant - no one will want to move wholesale into a Meta OS experience, and Apple are increasingly looking like they'll need to rely on google for high performance AI stuff.
All of this protects google's search business a lot.
Don't get me wrong, on the small stuff google is happy to let their people use 10% time to come up with a cool app which they'll kill after a couple of years, but for their big bets, every single time they've gone after something they have a lot to show for it where it counts to them.
Yeah, and Google has cared deeply about AI as a long term play since before they were public. And have been continuously invested there over the long haul.
The small stuff that they kill is just that--small stuff that was never important to them strategically.
I mean, sure, don't heavily invest (your attention, time, business focus, whatever) in something that is likely to be small to Google, unless you want to learn from their prototypes, while they do.
But to pretend that Google isn't capable of sustained intense strategic focus is to ignore what's clearly visible.
When did Google ever pivot?
I haven't followed that closely, but Gemini seems like a pivot based on ChatGPT's market success
Google is leading in terms of fundamental technology, but not in terms of products
They had the LLambda chatbot before that, but I guess it was being de-emphasized, until ChatGPT came along
Social was a big pivot, though that wasn't really due to Pichai. That was while Larry Page was CEO and he argued for it hard. I can't say anyone could have known beforehand, but in retrospect, Google+ was poorly conceived and executed
---
I also believe the Nth Google chat app was based on WhatsApp success, but I can't remember the name now
Google Compute Engine was also following AWS success, after initially developling Google App Engine
>I haven't followed that closely, but Gemini seems like a pivot based on ChatGPT's market success
"AI" in it's current form is already a massive threat to Google's main business (I personally use Google only a fraction of what I used to), so this pivot is justified.
Is it really such a pivot when they invested a lot in AI already?
They bought DeepMind in 2014 and always showed of a ton of AI research.
https://killedbygoogle.com
None of these are pivots. The core business has always been the core business.
If you are defining "pivot" as "abandon all other lines of business", then no, none of the BigTechs have ever pivoted.
By more reasonable standards of "pivot", the big investment into Google Plus/Wave in the social media era seems to qualify. As does the billions spent building out Stadia's cloud gaming. Not to mention the billions invested in their abandoned VR efforts, and the ongoing investment into XR...
G+ was a significant effort that was abandoned.
I'd personally define that as Google hedging their bet's and being prepared in case they needed to truly pivot, and then giving up when it became clear that they wouldn't need to. Sort of like "Apple Intelligence" but committing to the bit, and actually building something that was novel, and useful to some people, who were disappointed when it went away.
Stadia was always clearly unimportant to Google, and I say that as a Stadia owner (who got to play some games, and then got refunds.) As was well reported at the time, closing it was immaterial to their financials. Just because spending hundreds of millions of dollars or even a few billion dollars is significant to you or I doesn't mean that this was ever part of their core business.
Regardless, the overall sentimentality on HN about Google Reader and endless other indisputably small projects says more about the lack of strategic focus from people here, than it says anything about Alphabet.
Well, "pivot" implies the core business has failed and you're like "oh shit, let's do X instead".
Stadia was just Google's New Coke, Apple's Mac Cube, or Microsoft's MSNBC (or maybe Zune.
When they can't sell ads anymore, they'll have to pivot.
> Well, "pivot" implies the core business has failed and you're like "oh shit, let's do X instead".
I mean, Facebook's core business hasn't actually failed yet either, but their massive investments in short-form video, VR/XR/Metaverse, blockchain, and AI are all because they see their moat crumbling and are desperately casting around for a new field to dominate.
Google feels pretty similar. They made a very successful gambit into streaming video, another into mobile, and a moderately successful one into cloud compute. Now the last half a dozen gambits have failed, and the end of the road is in sight for search revenue... so one of the next few investments better pay off (or else)
The link you posted has a great many very insignificant investments included in it, and nothing I've seen Google doing has felt quite like the desperation of Facebook in recent years.
I didn't really see it at first, but I think you are correct to point out that they kind of rhyme. However to me, I think the clear desperation of Facebook makes it feel rather different from what I've seen Google doing over the years. I'm not sure I agree that Google's core business is in jeopardy in the way that Facebook's aging social media platform is.
Social. User-facing AI.
Social: YouTube User facing AI: Gemini, Google photos, NotebookLM and plenty of products.
I suppose you could argue that Amazon does have one special thing going for it here, idle compute resources in AWS. However that is not the sort of thing that requires "AI talent" to make use of.
They also have made pretty big investments in cloud VMs with GPUs attached, so they are making money off the AI craze regardless
The evidence shows that there is no methodological moat for LLMS. The moat of the frontier folks is just compute. xAI went in months from nothing to competing with the top dogs. DeepSeek too. So why bother with splurging billions in talent when you can buy GPUs and energy instead and serve the compute needs of everyone?
Also Amazon is in another capital intensive business. Retail. Spending billions on dubious AWS moonshots vs just buying more widgets and placing them across the houses of US customers for even faster deliveries does not make sense.
A lot of C-suite people seem to have an idea that if they just throw enough compute at LLMs that AGI will eventually emerge, even though it's pretty clear at this point that LLMs are never going to lead to general intelligence. In their view it makes sense to invest massive amounts of capital because it's like a lottery ticket to being the future AGI company that dominates the world.
I recall Zuckerberg saying something about how there were early signs of AI "improving itself." I don't know what he was talking about but if he really believes that's true and that we're at the bottom of an exponential curve then Meta's rabid hiring and datacenter buildout makes sense.
In early 2023, I remember someone breathlessly explaining that there are signs that LLMs that are seemingly good at chess/checkers moves may have a rudimentary model of the board within them, somehow magically encoded into the model weights through the training. I was stupid enough to briefly entertain the possibility until I actually bothered to develop a high level understanding of the transformer architecture. It's surprising how much mysticism this field seems to attract. Perhaps it being a non-deterministic, linguistically invoked black box, triggers the same internal impulses that draw some people to magic and spellcasting.
Just because it's not that hard to reach a high-level understanding of the transformer pipeline doesn't mean we understand how these systems function, or that there can be no form of world model that they are developing. Recently there has been more evidence for that particular idea [1]. The feats of apparent intelligence LLMs sometimes display have taken even their creators by surprise. Sure, there's a lot of hype too, that's part and parcel of any new technology today, but we are far from understanding what makes them perform so well. In that sense, yeah you could say they are a bit "magical".
[1] https://the-decoder.com/new-othello-experiment-supports-the-...
> Just because it's not that hard to reach a high-level understanding of the transformer pipeline doesn't mean we understand how these systems function
Mumbo jumbo magical thinking.
They perform so well because they are trained on probabilistic token matching.
Where they perform terribly, e.g math, reasoning, they are delegating to other approaches, and that's how you get the illusion that there is actually something there. But it's not. Faking intelligence is not intelligence. It's just text generation.
> In that sense, yeah you could say they are a bit "magical"
Nobody but the most unhinged hype pushers are calling it "magical". The LLM can never ever be AGI. Guessing the next word is not intelligence.
> there can be no form of world model that they are developing
Kind of impossible to form a world model if your foundation is probabilistic token guessing which is what LLMs are. LLMs are a dead end in achieving "intelligence", something novel as an approach needs to be discovered (or not) to go into the intelligence direction. But hey, at least we can generate text fast now!
> LLMs are a dead end in achieving "intelligence"
There is no evidence to indicate this is the case. To the contrary, all evidence we have points to these models, over time, being able to perform a wider range of tasks at a higher rate of success. Whether it's GPQA, ARC-AGI or tool usage.
> they are delegating to other approaches > Faking intelligence is not intelligence. It's just text generation.
It seems like you know something about what intelligence actually is that you're not sharing. If it walks, talks and quacks like a duck, I have to assume it's a duck[1]. Though, maybe it quacks a bit weird.
[1] https://en.wikipedia.org/wiki/Solipsism
> There is no evidence to indicate this is the case
Burden of proof is on those trying to convince us to buy into the idea of LLMs as being "intelligence".
There is no evidence of the Flying Spaghetti monster or Zeus or God not existing either, but we don't take seriously the people who claim they do exist (and there isn't proof because these concepts are made up).
Why should we take seriously the tolks claiming LLMs are intelligence without proof (there can't be proof, of course, because LLMs are not intelligence)?
Is there something we are all missing? Using Claude feels like magic sometimes, but can't everyone see the limitation now that we are 4 years and 100s of billions down the road?
Are they still really hoping that they are gonna tweak a model and feed it an even bigger dataset and it will be AGI?
I'm not a fan of mysticism. I'm also with you that these are simply statistical machines. But I don't understand what happened when understood transformers at a high-level.
If you're saying the magic disappeared after looking at a single transformer, did the magic of human intelligence disappear after you understood human neurons at a high level?
Its insane really, anyone who has worked with LLMs for a bit and has an idea of how they work shouldn't think its going to lead to "AGI".
Hopefully some big players, like FB bankrupt themselves.
Tbh I find this view odd, and I wonder what people view as agi now. It used to be that we had extremely narrow pieces of AI and I remember being on a research project about architectures and just very basic “what’s going on?” was advanced. Understanding that someone asked a question, that would be solved by getting a book and being able to then go and navigate to the place the book was likely to be was fancy. Most systems could solve literally one type of problem. They weren’t just bad at other things they were fundamentally incapable of anything but an extremely narrow use case.
I can throw wide ranging problems at things like gpt5 and get what seem like dramatically better answers than if I asked a random person. The amount of common sense is so far beyond what we had it’s hard to express. It used to be always pointed out that the things we had were below basic insect level. Now I have something that can research a charity, find grants and make coherent arguments for them, read matrix specs and debug error messages, and understand sarcasm.
To me, it’s clear that agi is here. But then what I always pictured from it may be very different to you. What’s your image of it?
It's more that "random" people are dumb as bricks (but we've in the name of equality and historic measurement errors decided to forgo that), add to it that AI's have a phenomenal (internet sized) memory makes them far more capable than many people.
However, even "dumb" people can often make judgements structures in a way that AI's cannot, it's just that many have such a bad knowledge-base that they cannot build the structures coherently whereas AI's succeed thanks to their knowledge.
I wouldn't be surprised if the top AI firms today spend an inordinate amount of time to build "manual" appendages into the LLM systems to cater to tasks such as debugging to uphold the facade that the system is really smart, while in reality it's mostly papering up a leaky model to avoid losing the enormous investments they need to stay alive with a hope that someone on their staff comes up a real solution to self-learning.
https://magazine.sebastianraschka.com/p/understanding-reason...
I think the discrepancy between different views on the matter mainly stems from the fact that state-of-the-art LLMs are better (sometimes extremely better) at some tasks, and worse (sometimes extremely worse) at other tasks, compared to average humans. For example, they're better at retrieving information from huge amounts of unstructured data. But they're also terrible at learning: any "experience" which falls out of the context window is lost forever, and the model can't learn from its mistakes. To actually make it learn something requires very many examples and a lot of compute, whereas a human can permanently learn from a single example.
> human can permanently learn from a single example
This, to me at least, seems like an important ingredient to satisfying a practical definition / implementation of AGI.
Another might be curiosity, and I think perhaps also agency.
I think it's clear that nobody agrees what AGI is. OpenAI describes it in terms of revenue. Other people/orgs in terms of, essentially, magic.
If I had to pick a name, I'd probably describe ChatGPT & co as advanced proof of concepts for general purpose agents, rather than AGI.
> I think it's clear that nobody agrees what AGI is
People selling AI products are incentivized to push misleading definitions of AGI.
Human-level intelligence. Being able to know what it doesn't know. Having a practical grasp on the idea of truth. Doing math correctly, every time.
I give it a high-res photo of a kitchen and ask it to calculate the volume of a pot in the image.
You discover truth by doing stuff in real world and observing the results. Current LLM have enough intelligence, but all the inputs they have are the “he said she said” by us monkeys, including all omissions and biases.
But many humans can't do a lot of those things and we still consider them "generally intelligent"
None of what you describe would I label within the realm of 'average'
It's not about what the average human can do - it's about what humans as a category are capable of. There will always be outliers (in both directions), but you can, in general, teach a human a variety of tasks, such as performing arithmetic deterministically, that you cannot teach to, for example, a parrot.
I don’t have a very high expectation of AGI at all. Just an algorithm or system you can put onto a robot dog, and get a dog level general intelligence. You should be able to live with that robot dog for 10 years and it should be just as capable as a dog throughout that timespan.
Hell, I’d even say we have AGI if you could emulate something like a hamster.
LLMs are way more impressive in certain ways than such a hypothetical AGI. But that has been true of computers for a long time. Computers have been much better at Chess than humans for decades. Dogs can’t do that. But that doesn’t mean that a chess engine is an AGI.
I would also say we have a special form of AGI if the AI can pass an extended Turing test. We’ve had chat bots that can fool a human for a minute for a long time. Doesn’t mean we had AGI. So time and knowledge was always a factor in a realistic Turing test. If an AGI can fool someone who knows how to properly probe an LLM, for a month or so, while solving a bunch of different real world tasks that require stable long term memory and planning, then I’d day we’re in AGI territory for language specifically. I think we have to distinguish between language AGI and multi-modal AGI. So this test wouldn’t prove what we could call “full” AGI.
These are some of the missing components for full AGI: - Being able to act as a stable agent with a stable personality over long timespans - Capable of dealing with uncertainties. Having a understanding of what it doesn’t know - One-shot learning, with long term retention, for a large number of things - Fully integrated multi-modality across sound, vision, and other inputs/outputs we may throw at it.
The last one is where we may be able to get at the root of the algorithm we’re missing. A blind person can learn to “see” by making clicks and using their ears to see. Animals can do similar “tricks”. I think this is where we truly see the full extent of the generality and adaptability of the biological brain. Imagine trying to make a robot that can exhibit this kind of adaptability. It doesn’t fit into the model we have for AI right now.
my picture of AGI is 1) autonomous improvement 2) ability to say 'i don't know/can't be done'
I wonder if 2) is a result of published bias for positive results in the training set. An “I don’t know” response is probably ranked unsatisfactory by human feedback and most published scientific literature are biased towards positive results and factual explanations.
In my experience, the willingness to say "I don't know" instead of confabulate is also down-rated as a human attribute, so it's not surprising that even an AGI trained on the "best" of humanity would avoid it.
Nobody is saying that LLM's don't work like magic. I know how neural networks work and they still feel like voodoo to me.
What we are saying is that LLM's can't become AGI. I don't know what AGI will look like, but it won't look like an LLM.
There is a difference between being able to melt iron and being able to melt tungsten.
Recently I realized that US are very close to a centrally planned economy. Meta wasted 50B on metaverse, which like how much Texas spends on healthcare. Now the "AI" investments seems dubious.
You could fund 1000+ projects with this kinds of money. This is not an effective capital allocation.
The only way we'll have AGI is if people get dumber. Using modern tech like LLMs makes people dumber. Ergo, we might see AGI sooner than expected.
> ... and has an idea of how they work shouldn't think its going to lead to "AGI"
Not sure what level of understanding are you referring to but having learned and researched about the pretty much all LLM internals I think this has led me exactly to the opposite line of thinking. To me it's unbelievable what we have today.
I think AI research is like anything else really. The smartest people are heads down working on their problems. The people going on podcasts are less connected to day to day work.
It’s also pretty useless to talk about whether something is AGI without defining intelligence in the first place.
I think something like we saw in the show "Devs" is much more likely, although what the developers did with it in the show was bonkers unrealistic. But some kind of big enough quantum device basically.
Just scaling them up might not leat to "AGI", but they can still lead to AGI as a bridge.
This is not and has not been the consensus opinion. If you're not an AI researcher you shouldn't write as if you've set your confidence parameter to 0.95.
Of course it might be the case, but it's not a thing that should be expressed with such confidence.
Is it widely accepted that LLMs won't lead to AGI? I've asked Gemini, so it came up with four primary arguments for this claim, commenting on them briefly:
1) LLMs as simple "next token predictors" so they just mimicry thinking: But can it be argued that current models operate on layers of multiple depth and are able to actually understand by building concepts and making connections on abstract levels? Also, don't we all mimicry?
2) Grounding problem: Yes, models build their world models on text data, but we have models operating on non-textual data already, so this appears to be a technical obstacle rather than fundamental.
3) Lack of World Model. But can anyone really claim they have a coherent model of reality? There are flat-earthers, yet I still wouldn't deny them having AGI. People hallucinate and make mistakes all the time. I'd argue hallucinations is in fact the sign of an emerging intelligence.
4) Fixed learning data sets. Looks like this is now being actively solved with self-improving models?
I just couldn't find a strong argument supporting this claim. What am I missing?
Why on earth would you copy and paste an LLM's output into a comment? What does that accomplish or provide that just a simply stated argument doesn't accomplish more succinctly? If you don't know something, simply don't comment on it -- or ask a question.
None of the above is AI.
> I've asked Gemini, so it came up with four primary arguments for this claim, commenting on them briefly:
This line means, and literally says, that everything that follows is a summary or direct quotation from an LLM's output.
There's a more charitable but unintuitive interpretation, in which "commenting on them briefly" is intended to mean "I will comment on them briefly:". But this isn't a natural interpretation. It's one I could be expected to reach only after seeing your statement that 'none of the above is AI.' But even this more charitable interpretation actually contradicts your claim that it's not AI.
So now I'm even less sure I know what you meant to communicate. Either I'm missing something really obvious or the writing doesn't communicate what you intended.
Fur future reference, pasting llm slop feels exactly as patronizing as back when people pasted links to google searches in response to questions they considered beneath their dignity to answer. Except in this case, no-one asked to begin with.
> I don't know what he was talking about
There's a bunch of ways AI is improving itself, depending on how you want to interpret that. But it's been true since the start.
1. AI is used to train AI. RLHF uses this, curriculum learning is full of it, video model training pipelines are overflowing with it. AI gets used in pipelines to clean and upgrade training data a lot.
2. There are experimental AI agents that can patch their own code and explore a tree of possibilities to boost their own performance. However, at the moment they tap out after getting about as good as open source agents, but before they're as good as proprietary agents. There isn't exponential growth. There might be if you throw enough compute at it, but this tactic is very compute hungry. At current prices it's cheaper to pay an AI expert to implement your agent than use this.
So have an AI with a 40% error rate judge an AI with an 40% error rate…
AGI is a complete no go until a model can adjust its own weights on the fly, which requires some kind of negative feedback loop, which requires a means to determine a failure.
Humans have pain receptors to provide negative feedback and we can imagine events that would be painful such as driving into a parked car would be painful without having to experience it.
If current models could adjust its own weights to fix the famous “how many r’s in strawberry” then I would say we are on the right path.
However, the current solution is to detect the question and forward it to a function to determine the right answer. Or attempt to add more training data the next time the model is generated ($$$). Aka cheat the test.
I think LLM as a toolsmith like demonstrated in the Voyager paper (1) is another interesting approach to creating a system that can learn to do a task better over time. (1) https://arxiv.org/abs/2305.16291
> There are experimental AI agents that can patch their own code and explore a tree of possibilities to boost their own performance. However, at the moment they tap out after getting about as good as open source agents, but before they're as good as proprietary agents.
Interesting. Do you have links?
Not OP, but the Darwin Godel machine comes to mind: https://arxiv.org/abs/2505.22954
That's the one!
Thanks!
I'm skeptical that RLHF really works. Doesn't it just patch the obvious holes so it looks better on paper? If it can't reason then it will continue to get 2nd and 3rd order difficulty problems wrong.
> it's pretty clear at this point that LLMs are never going to lead to general intelligence.
It is far from clear. There may well be emergent hierarchies of more abstract thought at much higher numbers of weights. We just don't know how a transformer will behave if one is built with 100T connections - something that would finally approach the connectome level of a human brain. Perhaps nothing interesting but we just do not know this and the current limitation in building such a beast is likely not software but hardware. At these scales the use of silicon transistors to approximate analog curve switching models just doesn't make sense. True neuromorphic chips may be needed to approach the numbers of weights necessary for general intelligence to emerge. I don't think there is anything in production at the moment that could rival the efficiency of biological neurons. Most likely we do not need that level of efficiency. But it's almost certain that stringing together a bunch of H100s isn't a path to the scale we should be aiming for.
I don't get it, I really don't.
Even assuming a company gets to AGI first this doesn't mean another one will follow.
Suppose that FooAI gets to it first: - competitors may get there too in a different or more efficient way - Some FooAI staff can leave and found their own company - Some FooAI staff can join a competitor - FooAI "secret sauce" can be figured out, or simply stolen, by a competitor
At the end of the day, it really doesn't matter, the equation AI === commodity just does not change.
There is no way to make money by going into this never ending frontier model war, price of training keeps getting higher and higher, but your competitors few months later can achieve your own results for a fraction of your $.
Some would say that the race to AGI is like the race to nuclear weapons and that the first to get there will hold all the cards (and be potentially able to stop others getting there.) It's a bit too sci-fi for me.
If AGI is reached it would be trivial for the competing superpowers to completely quarantine themselves network wise by cutting undersea cables long enough to develop competing AGI
I don't know if AGI will emerge from LLM, but I'm always reminded of the Chinese room thought experiment. With billions thrown at the idea it will certainly be the ultimate answer as to whether true understanding can emerge from a large enough dictionary.
Please stop refering to the Chinese Room - it's just magical/deist thinking in disguise. It postulates that humans have way of 'understanding' things that is impossible to replicate mechanically.
The fact that philosophy hasn't recognized and rejected this argument based on this speaks volumes of the quality of arguments accepted there.
(That doesn't mean LLMs are or will be AGI, its just this argument is tautological and meaningless)
That some people use the Chinese Room to ascribe some magical properties to human consciousness says more about the person drawing that conclusion than the thought experiment itself.
I think it's entirely valid to question whether a computer can form an understanding through deterministically processing instructions, whether that be through programming language or language training data.
If the answer is no, that shouldn't lead to a deist conclusion. It can just as easily lead to the conclusion that a non-deterministic Turing machine is required.
I'd appreciate if you tried to explain why instead of resorting to ad hominem.
> I think it's entirely valid to question whether a computer can form an understanding through deterministically processing instructions, whether that be through programming language or language training data.
Since the real world (including probabilistic and quantum phenomena) can be modeled with deterministic computation (a pseudorandom sequence is deterministic, yet simulates randomness), if we have a powerful enough computer we can simulate the brain to a sufficient degree to have it behave identically as the real thing.
The original 'Chinese Room' experiment describes a book of static rules of Chinese - which is probably not the way to go, and AI does not work like that. It's probabilistic in its training and evaluation.
What you are arguing is that constructing an artificial consciousness lies beyond our current computational ability(probably), and understanding of physics (possibly), but that does not rule out that we might solve these issues at some point, and there's no fundamental roadblock to artificial consciousness.
I've re-read the argument (https://en.wikipedia.org/wiki/Chinese_room) and I cannot help but conclude that Searle argues that 'understanding' is only something that humans can do, which means that real humans are special in some way a simulation of human-shaped atoms are not.
Which is an argument for the existence of the supernatural and deist thinking.
> I'd appreciate if you tried to explain why instead of resorting to ad hominem.
It is not meant as an ad hominem. If someone thinks our current computers can't emulate human thinking and draws the conclusion that therefore humans have special powers given to them by a deity, then that probably means that person is quite religious.
I'm not saying you personally believe that and therefore your arguments are invalid.
> Since the real world (including probabilistic and quantum phenomena) can be modeled with deterministic computation (a pseudorandom sequence is deterministic, yet simulates randomness), if we have a powerful enough computer we can simulate the brain to a sufficient degree to have it behave identically as the real thing.
The idea that a sufficiently complex pseudo-random number generator can emulate real-world non-determinism enough to fully simulate the human brain is quite an assumption. It could be true, but it's not something I would accept as a matter of fact.
> I've re-read the argument (https://en.wikipedia.org/wiki/Chinese_room) and I cannot help but conclude that Searle argues that 'understanding' is only something that humans can do, which means that real humans are special in some way a simulation of human-shaped atoms are not.
In that same Wikipedia article Searle denies he's arguing for that. And even if he did secretly believe that, it doesn't really matter, because we can draw our own conclusions.
Disregarding his arguments because you feel he holds a hidden agenda, isn't that itself an ad hominem?
(Also, I apologize for using two accounts, I'm not attempting to sock puppet)
What are his arguments then?
>Searle argues that, without "understanding" (or "intentionality"), we cannot describe what the machine is doing as "thinking" and, since it does not think, it does not have a "mind" in the normal sense of the word.
This is the only sentence that seems to be pointing to what constitutes the specialness of humans, and the terms of 'understanding' and 'intentionality' are in air quotes so who knows? This sounds like the archetypical no true scotsman fallacy.
In mathematical analysis, if we conclude that the difference between 2 numbers is smaller than any arbitrary number we can pick, those 2 numbers must be the same. In engineering, we can reduce the claim to 'any difference large about to care about'
Likewise if the difference between a real human brain and an arbitrarily sophisticated Chinese Room brain is arbitrarily small, they are the same.
If our limited understanding of physics and engineering makes the practical difference not zero, this essentially becomes a bit of a somewhat magical 'superscience' argument claiming we can't simulate the real world to a good enough resolution that the meaningful differences between our 'consciousness simulator' and the thing itself disappear - which is an extraordinary claim.
> What are his arguments then?
They're in the "Complete Argument" section of the article.
> This sounds like the archetypical no true scotsman fallacy.
I get what you're trying to say, but he is not arguing only a true Scotsman is capable of thought. He is arguing that our current machines lack the required "causal powers" for thought. Powers that he doesn't prescribe to only a true Scotsman, though maybe we should try adding bagpipes to our AI just to be sure...
Thanks, but that makes his arguments even less valid.
He argues that computer programs only manipulate symbols and thus have no semantic understanding.
But that's not true - many programs, like compilers that existed back when the argument was made, had semantic understanding of the code (in a limited way, but they did have some understanding about what the program did).
LLMs in contrast have a very rich semantic understanding of the text they parse - their tensor representations encode a lot about each token, or you can just ask them about anything - they might not be human level at reading subtext, but they're not horrible either.
Now you're getting to the heart of the thought experiment. Because does it really understand the code or subtext, or is it just really good at fooling us that it does?
When it makes a mistake, did it just have a too limited understanding or did it simply not get lucky with its prediction of the next word? Is there even a difference between the two?
I would like to agree with you that there's no special "causal power" that Turing machines can't emulate. But I remain skeptical, not out of chauvinism, but out of caution. Because I think it's dangerous to assume an AI understands a problem simply because it said the right words.
> I cannot help but conclude that Searle argues that ‘understanding’ is only something that humans can do, which means…
Regardless of whether Searle is right or wrong, you’ve jumped to conclusions and are misunderstanding his argument and making further assumptions based on your misunderstanding. Your argument is also ad-hominem by accusing people of believing things they don’t believe. Maybe it would be prudent to read some of the good critiques of Searle before trying to litigate it rapidly and sloppily on HN.
The randomness stuff is very straw man, definitely not a good argument, best to drop it. Today’s LLMs are deterministic, not random. Pseudorandom sequences come in different varieties, but they model some properties of randomness, not all of them. The functioning of today’s neural networks, both training and inference, is exactly a book of static rules, despite their use of pseudorandom sequences.
In case you missed it in the WP article, most of the field of cognitive science thinks Searle is wrong. However, they’re largely not critiquing him for using metaphysics, because that’s not his argument. He’s arguing that biology has mechanisms that binary electronic circuitry doesn’t; not human brains, simply physical chemical and biological processes. That much is certainly true. Whether there’s a difference in theory is unproven. But today currently there absolutely is a difference in practice, nobody has ever simulated the real world or a human brain using deterministic computation.
If scientific consensus is that he's wrong why is he being constantly brought up and defended - am I not right to call them out then?
Nobody brings up that light travels through the aether, that diseases are caused by bad humors etc. - is it not right to call out people for stating theory that's believed to be false?
>The randomness stuff is very straw man,
And a direct response to what armada651 wrote:
>I think it's entirely valid to question whether a computer can form an understanding through deterministically processing instructions, whether that be through programming language or language training data.
> He’s arguing that biology has mechanisms that binary electronic circuitry doesn’t; not human brains, simply physical chemical and biological processes.
Once again the argument here changed from 'computers which only manipulate symbols cannot create consciousness' to 'we don't have the algorithm for consiousness yet'.
He might have successfully argued against the expert systems of his time - and true, mechanistic attempts at language translation have largely failed - but that doesn't extend to modern LLMs (and pre LLM AI) or even statistical methods.
You’re making more assumptions. There’s no “scientific consensus” that he’s wrong, there are just opinions. Unlike the straw man examples you bring up, nobody has proven the claims you’re making. If they had, then the argument would go away like the others you mentioned.
Where did the argument change? Searle’s argument that you quoted is not arguing that we don’t have the algorithm yet. He’s arguing that the algorithm doesn’t run on electrical computers.
I’m not defending his argument, just pointing out that yours isn’t compelling because you don't seem to fully understand his, at least your restatement of it isn’t a good faith interpretation. Make his argument the strongest possible argument, and then show why it doesn’t work.
IMO modern LLMs don’t prove anything here. They don’t understand anything. LLMs aren’t evidence that computers can successfully think, they only prove that humans are prone to either anthropomorphic hyperbole, or to gullibility. That doesn’t mean computers can’t think, but I don’t think we’ve seen it yet, and I’m certainly not alone there.
>most of the field of cognitive science thinks Searle is wrong.
>There’s no “scientific consensus” that he’s wrong, there are just opinions.
And? Are you imagining that these aren’t both true at the same time? If so, I’m happy to explain. Since nothing has been proven, there’s nothing “scientific”. And since there’s some disagreement, “consensus” has not been achieved yet. This is why your presumptive use of “scientific consensus” was not correct, and why the term “scientific consensus” is not the same thing as “most people think”. A split of 60/40 or 75/25 or even 90/10 counts as “most” but does not count as “consensus”. So I guess maybe be careful about assuming what something means, it seems like this thread was limited by several incorrect assumptions.
> The fact that philosophy hasn't recognized and rejected this argument based on this speaks volumes of the quality of arguments accepted there.
That's one possibility. The other is that your pomposity and dismissiveness towards the entire field of philosophy speaks volumes on how little you know about either philosophical arguments in general or this philosophical argument in particular.
Another ad hominem, I'd like you to refute my claim that Searle's argument is essentially 100% magical thinking.
And yes, if for example, medicine would be no worse at curing cancer than it is today, yet doctors asserted that crystal healing is a serious study, that would reflect badly on the field at large, despite most of it being sound.
Searle refutes your claim that there’s magical thinking.
“Searle does not disagree with the notion that machines can have consciousness and understanding, because, as he writes, "we are precisely such machines". Searle holds that the brain is, in fact, a machine, but that the brain gives rise to consciousness and understanding using specific machinery.”
But the core of the original argument is that programs only manipulate symbols and consciousness can never arise just through symbol manipulation - which here then becomes 'we have not discovered the algorithms' for consciousness yet.
It's just a contradiction.
When you say something that contradicts his statements, it doesn’t mean he’s wrong, it most likely means you haven’t understood or interpreted his argument correctly. The Wikipedia page you linked to doesn’t use the word “algorithm”, so the source of the contradiction you imagine might be you. Searle says he thinks humans are biological machines, so your argument should withstand that hypothesis rather than dismiss it.
Why on earth do you take it as an ad hominem attack? Do you really think your comment isn't dismissive or pompous?
Another ad hominem, just like you calling anyone who talks about the chinese room thought experiment a deist?
It is still relevant because it hasn’t been disproven yet. So far all computer programs are Chinese Rooms, LLM’s included.
If you’re talking about it being proven or disproven you’re misunderstanding the point of the thought experiment.
"Please stop referring to this thought experiment because it has possible interpretations I don't personally agree with"
Please give me an interpretation that is both correct an meaningful (as in possible to disprove)
The human way of understanding things can be replicated mechanically, because it is mechanical in nature. The contents of your skull are an existence proof of AGI.
The A stands for artificial.
The contents of my skull are only a proof for AGI if your mechanical machine replicates all its processes. It's not a question about whether a machine can reproduce that, it's a question about whether we have given our current machines all the tools it needs to do that.
The theory of special relativity does not say 'you can't exceed the speed of light(unless you have a really big rocket)'. It presents a theoretical limit. Likewise the Chinese room doesn't state that consciousness is an intractable engineering problem, but an impossibility.
But the way Searle formulates his argument, by not defining what consciousness is, he essentially gives himself enough wiggle room to be always right - he's essentially making the 'No True Scotsman' fallacy.
The moat is people, data, and compute in that order.
It’s not just compute. That has mostly plateaued. What matters now is quality of data and what type of experiments to run, which environments to build.
This "moat" is actually constantly shifting (which is why it isn't really a moat to begin with). Originally, it was all about quality data sources. But that saturated quite some time ago (at least for text). Before RLHF/RLAIF it was primarily a race who could throw more compute at a model and train longer on the same data. Then it was who could come up with the best RL approach. Now we're back to who can throw more compute at it since everyone is once again doing pretty much the same thing. With reasoning we now also opened a second avenue where it's all about who can throw more compute at it during runtime and not just while training. So in the end, it's mostly about compute. The last years have taught us that any significant algorithmic improvement will soon permeate across the entire field, no matter who originally invented it. So people are important for finding this stuff, but not for making the most of it. On top of that, I think we are very close to the point where LLMs can compete with humans on their own algorithmic development. Then it will be even more about who can spend more compute, because there will be tons of ideas to evaluate.
To put that into a scientific context - compute is capacity to do experiments and generate data ( about how best to build models ).
However I do think you are missing an important aspect - and that's people who properly understand important solvable problems.
ie I see quite a bit "we will solve this x, with AI' from startup's that don't fundamentally understand x.
>we will solve this x, with AI
You usually see this from startup techbro CEOs understand neither x nor AI. Those people are already replacable by AI today. The kind of people who think they can query ChatGPT once with "How to create a cutting edge model" and make millions. But when you go in on the deep end, there are very few people who still have enough tech knowledge to compete with your average modern LLM. And even the Math Olympiad gold medalists high-flyers at DeepSeek are about to have a run for their money with the next generation. Current AI engineers will shift more and more towards senior architecture and PM roles, because those will be the only ones that matter. But PM and architecture is already something that you could replace today.
> Originally, it was all about quality data sources.
It still is! Lots of vertical productivity data that would be expensive to acquire manually via humans will be captured by building vertical AI products. Think lawyers, doctors, engineers.
That's literally what RLAIF has been doing for a while now.
People matter less and less as well.
As more opens up in OSS and academic space, their knowledge and experience will either be shared, rediscovered, or become obsolete.
Also many of the people are coasting on one or two key discoveries by a handful of people years ago. When Zuck figures this out he gonna be so mad.
Not all will become OSS. Some will become products, and that requires the best people.
Lets not pretend this is strategy. Amazon has been trying and failing to hire top AI people. No-one in their right minds would join. Even Meta has to shell out 8-9 figures for top people, who with any modicum of talent or self respect would go to Amazon rather than Anthropic, OAI, GDM? They bought Adept, everyone left.
AWS is also falling far behind Azure wrt serving AI needs at the frontier. GCP is also growing at a faster rate and has a way more promising future than AWS in this space.
AWS is very far behind, its already impacting the stock. Without a winning AI offering, all new cloud money is going to GCP and Azure. They have a huge problem
> The evidence shows that there is no methodological moat for LLMS.
Does it? Then how come Meta hasn't been able to release a SOTA model? It's not for a lack of trying. Or compute. And it's not like DeepSeek had access to vastly more compute than other Chinese AI companies. Alibaba and Baidu have been working on AI for a long time and have way more money and compute, but they haven't been able to do what DeepSeek did.
They may not have been leading (as in, releasing a SOTA model), but they definitely can match others - easily, as shown by llama 3/4, which proves the point - there is no moat. With enough money and resources, you can match others. Whether without SOTA models you can make a business out of it is a different question.
Meta never matched the competition with their Llama models. They've never even come close. And Llama 4 was an actual disaster.
I am not a daily user, so only rely on reviews and benchmarks - actual experience may be different.
Even in reviews and benchmark, Llama wasn't close to frontier models. Also Llama 2/3 lead in open weight models wasn't more than few months.
> but they definitely can match others - easily, as shown by llama 3/4
Are we living in the same universe? LLAMA is universally recognized as one of the worst and least successful model releases. I am almost certain you haven't ever tried a LLAMA chat, because, by the beard of Thor, it's the worst experience anyone could ever had, with any LLM.
LLAMA 4 (behemoth, whatever, whatever) is an absolute steaming pile of trash, not even close to ChatGPT 4o/4/5/, Gemini(any) and even not even close to cheaper ones like DeepSeek. And to think Meta pirated torrents to train it...
What a bunch of criminal losers and what a bunch of waste of money, time and compute. Oh, at least the Metaverse is a success...
https://www.pcgamer.com/gaming-industry/court-documents-show...
https://www.cnbc.com/2025/06/27/the-metaverse-as-we-knew-it-...
> The moat of the frontier folks is just compute.
This is not really true. Google has all the compute but in many dimensions they lag behind GPT-5 class (catching up, but it has not been a given).
Amazon itself did try to train a model (so did Meta) and had limited success.
I switched to Gemini with my new phone and I literally couldn't tell a difference. It is actually crazy how small the cost of switching is for LLMs. It feels like AI is more like a commodity than a service.
> I switched to Gemini with my new phone and I literally couldn't tell a difference. It is actually crazy how small the cost of switching is for LLMs. It feels like AI is more like a commodity than a service.
It is. It's wild to me that all these VCs pouring money into AI companies don't know what a value-chain is.
Tokens are the bottom of the value-chain; it's where the lowest margins exist because the product at that level is a widely available commodity.
I wrote about this already (shameless plug: https://www.rundata.co.za/blog/index.html?the-ai-value-chain )
On top of that, the on-device models have got stronger and stronger as the base models + RL has got better. You can do on your laptop now what 2 years ago was state of the art.
Which dimensions do you see Google lagging on? They seem broadly comparable on the usual leaderboard (https://lmarena.ai/leaderboard) and anecdotally I can't tell the difference in quality.
I tend personally to stick with ChatGPT most of the time, but only because I prefer the "tone" of the thing somehow. If you forced me to move to Gemini tomorrow I wouldn't be particularly upset.
> Which dimensions do you see Google lagging on? They seem broadly comparable on the usual leaderboard (https://lmarena.ai/leaderboard) and anecdotally I can't tell the difference in quality.
Gemini holds indeed the top spot, but I feel you framed your response quite well: they are all broadly comparable. The difference in the synthetic benchmark from the top spot and the 20th spot was something like 57 points on a scale of 0-1500
" in many dimensions they lag behind GPT-5 class " - such as?
Outside of computer, "the moat" is also data to train on. That's an even wider moat. Now, google has all the data. Data no one else has or ever will have. If anything, I'd expect them to outclass everyone by a fat margin. I think we're seeing that on video however.
You think Chinese companies are short on data and people? Google doesn’t have an advantage there until the CCP takes on a more hands on approach.
Tin foil hat time:
- If you were a God and you wanted to create an ideal situation for the arrival of AI
- It would make sense to precede it with a social media phenomena that introduces mass scale normalization of sharing of personal information
Yes, that would be ideal …
People can’t stop sharing and creating data on anything, for awhile now. It’s a perfect situation for AI as an independent, uncontrollable force.
> People can’t stop sharing and creating data on anything
Garbage in. Garbage out.
There has never been a better time to produce an AI that mimics a racist uneducated teenager.
We already had Tay. https://en.wikipedia.org/wiki/Tay_(chatbot)
Do you want to model the world accurately or not? That person is part of our authentic reality. The most sophisticated AI in the world will always include that person(s).
Not in the slightest. I want useful information services that behave in a mature and respectable fashion.
not according to google: “We have no moat, and neither does OpenAI”: the big memo and the big HN thread on same https://news.ycombinator.com/item?id=35813322
a bit weird to think about it since google has literally internet.zip in multiple versions over the years, all of email, all of usenet, all of the videos, all of the music, all of the user's search interest, ads, everything..
> a bit weird to think about it since google has literally internet.zip in multiple versions over the years, all of email, all of usenet, all of the videos, all of the music, all of the user's search interest, ads, everything..
Yeah, Google totally has a moat. Them saying that they have no moat doesn't magically make that moat go away.
They also own the entire vertical which none of the competitors do - all their competitors have to buy compute from someone who makes a profit just on compute (Nvidia, for example). Google owns the entire vertical, from silicon to end-user.
It would be crazy if they can't make this work.
That's why robots make so much traffic now. Those other companies are trying to get data.
Google theoretically has reddit access. I wonder if they have sort of an internet archive - data unpolutted by LLMs
On a side note, funny how all the companies seem to train on book archivr which they just downloaded from the internet
> all of the videos, [...], all of the user's search interest, ads, everything..
And privacy policies that are actually limiting what information gets used in what.
and even then!
That's one person's opinion that works for Google.
counterpoint: with their aggressive crawlers, most AI companies can have as much data as google...
> This is not really true. Google has all the compute but in many dimensions they lag behind GPT-5 class (catching up, but it has not been a given).
I don't know what you are talking about. I use Gemini on a daily basis and I honestly can't tell a difference.
We are at a point where training corpus and hallucinations makes more of a difference than "model class".
Yes, or Apple who with all the talent don’t manage to pull off anything useful in AI
xAI seems to be the exception, not the rule
Given Apple’s moat is their devices, their particular spin on AI is very much edge focussed, which isn’t as spectacular as the current wave of cloud based LLM. Apple’s cloud stuff is laughably poor.
Depending on how you look at it I suppose but I believe Gemini surpasses OpenAI on many levels now. Better photo and video models. The leaderboard for text and embeddings are also putting Google on top of Openai.
gemini-2.5-pro is ranked number 1 in llmarena (https://lmarena.ai/leaderboard) before gpt-5-high. In the Text-to-Video and Image-to-video, google also have the highest places, OpenAI is nowhere.
Yes, but they're also slower. As LLMs start to be used for more general purpose things, they are becoming a productivity bottle-neck. If I get a mostly right answer in a few seconds that's much better than a perfect answer in 5 minutes.
Right now the delay for Google's AI coding assistant is high enough for humans to context switch and do something else while waiting. Particularly since one of the main features of AI code assistants is rapid iteration.
Anecdotally, Gemini pro is way faster than GPT 5 thinking. Flash is even faster. I have no numbers though.
It doesn’t guarantee success, but the point stands about X and Deepseek
I think I'm right in saying that AWS, rather than deliveries, is by far the most profitable part of Amazon.
Also a smart move is to be selling shovels in a gold rush - and that's exactly what Amazon is doing with AWS.
Amazon retail runs on ridiculously low margins compared to AWS. Revenue-wise retail dwarfs AWS, profit-wise it's vice-versa.
The barriers to entry for LLM are obvious: as you pointed, the field is extremely capital intensive. The only reason there are seemingly multiple players is because the amount of capital thrown at it at the moment is tremendous but that's unlikely to last forever.
From my admittely poorly informed point of view, strategy-wise, it's hard to tell how wise it is investing in foundational work at the moment. As long as some players release competitive open weight models, the competitive advantage of being a leader in R&D will be limited.
Amazon already has the compute power to place itself as a reseller without investing or having to share the revenue generated. Sure, they won't be at the forefront but they can still get their slice of the pie without exposing themselves too much to an eventual downturn.
The idea that models are copyrightable is also extremely dubious.
So there probably isn’t even a legal moat.
There's not much of an architectural moat, but there is a methodological moat, such as with RL synthetic data.
Are you arguing anthropic has more compute than Amazon?
Are you saying the only reason Meta is behind everyone else is compute????
Think well: why should a platform provider get into a terribly expensive and unprofitable business when they can just provide hardware for those with money to spend? This was AWS strategy for years and it's been working well for them.
> Are you arguing anthropic has more compute than Amazon?
I wouldn't be surprised if the likes of Anthropic wasn't paying AWS for its compute.
As the saying goes, the ones who got rich from the gold rush were the ones selling shovels.
I wouldn't be surprised if Amazon just buys Anthropic or another lab rather than competing for individuals.
Does Amazon want to be an AI innovator or an AI enabler?
AWS enables thousands of other companies to run their business. Amazon has designed their own Graviton ARM CPUS and their own Trainium AI chips. You can access these through AWS for your business.
I think Amazon sees AI being used in AWS as a bigger money generator than designing new AI algorithms.
Also I think that they realize this is just a money losing proposition right now for the most part. And they're not going to have a problem getting in later when there's a clear solution. Why fight it out? I don't think they're going to miss much because they can use any models they need and as you said some of that stuff may be run on their servers
I can make a case: Building their own models like Nova and Titan allow them to build up expertise in how to solve hyperscaler problems. Think of it like Aurora, where they have a generally solved problem (RDBMS) but it needs to be modified to work with the existing low-level primitives. Yes, it can be done in the open, but if I'm AWS, I probably want to jealously guard anything that could be a key differentiator.
Reading comments from the appropriate VPs will illuminate the situation.. Swami is looking to democratise AI, and the company is geared towards that more than anything else.
Disclaimer; I work for amzn, opinions my own.
https://aws.amazon.com/blogs/machine-learning/aws-and-mistra...
I don't know what democratizing AI means, AWS doesn't have the GPU infrastructure to host inference or training on a large scale.
I started this part of the thread and mentioned Trainium but the person you replied to gave a link. Follow that and you can see Amazon's chips that they designed.
Amazon wants people to move away from Nvidia GPUs and to their own custom chips.
https://aws.amazon.com/ai/machine-learning/inferentia/
https://aws.amazon.com/ai/machine-learning/trainium/
TBH I was just going off of that I've heard AWS is a terrible place to get h100 clusters at scale. And for the training I was looking at we didn't really want to consider going off CUDA.
Huh? That’s quite the assertion. They provide the infrastructure for Anthropic, so if that’s not large scale idk what is.
They have to use GCP as well, which is arguably a strong indictment of their experience with AWS. Coincidentally, this aligns with my experience trying to train on AWS.
It’s unclear why Swami is put in charge of this stuff. He’s not a recognized leader in the space and hasn’t delivered a coherent strategy. However, per the article Amazon is struggling to hire and retain the best talent and thus it may just be the best they have.
Who is “Swami?” Although I suppose that’s just making the point that Amazon’s folks aren’t recognized leaders in this space.
Selling pick axes vs. mining for gold yet again!
I'm glad this analogy is at the top. I think that some large companies like AWS really should not try to blow money on AI in ways that only make a lot more sense for companies like Meta, Google, and Apple. AWS can't trap you in their AI systems with network effects that the other competitors can.
Companies like OpenAI and Anthropic are still incredibly risky investments especially because of the wild capital investments and complete lack of moat.
At least when Facebook was making OpenAI's revenue numbers off of 2 billion active users it was trapping people in a social network where there were real negative consequences to leaving. In the world of open source chatbots and VSClone forks there's zero friction to moving on to some other solution.
OpenAI is making $12 billion a year off of 700 million users [1], or around $17 per user annually. What other products that have no ad support perform that badly? And that's a company that is signing enterprise contracts with companies like Apple, not just some Spotify-like consumer service.
[1] This is almost the exact same user count that Facebook had when it turned its first profit.
> OpenAI is making $12 billion a year off of 700 million users [1], or around $17 per user annually. What other products that have no ad support perform that badly?
That's a bit of a strange spin. Their ARPU is low because they are choosing not to monetize 95% of their users at all, and for now are just providing practically limitless free service.
But monetising those free users via ads will pretty obviously be both practical and lucrative.
And even if there is no technical moat, they seem to have a very solid mind share moat for consumer apps. It isn't enough for competitors to just catch up. They need to be significantly better to shift consumer habits.
(For APIs, I agree there is no moat. Switching is just so easy.)
> They need to be significantly better to shift consumer habits.
i am hoping that a device local model would eventually be possible (may be a beefy home setup, and then an app that connects to your home on mobile devices for use on the go).
currently, hardware restrictions prevent this type of home setup (not to mention the open source/free models aren't quite there and difficulty for non-tech users to actually setup). However, i choose to believe the hardware issues will get solved, and it will merely be just time.
The software/model issue, on the other hand is harder to see solved. I pin my hopes onto deepseek, but may be meta or some other company will surprise me.
I think you're super wrong about the local model issue and that's a huge risk for companies like OpenAI.
Apple products as an example have an excellent architecture for local AI. Extremely high-bandwidth RAM.
If you run an OSS model like gpt-oss on a Mac with 32GB of RAM it's already very similar to a cloud experience.
i dont have the hardware to run or try them, but from the huggingfaces discussion forums, gpt-oss seems to be pretty hard censored. I would not consider it as being a viable self-hosted LLM except for the very narrowest of domains (like coding for example).
I'm not sure where censorship comes in with this discussion, it seems like cloud models are censored as well? And local models are frequently created that are abliterated? Correct me if I'm wrong or misunderstanding you.
Either way, it's just an example model, plenty of others to choose from. The fact of the matter is that the base model MacBook Air currently comes with about half as much RAM as you need for a really really decent LLM model. The integrated graphics are fast/efficient and the RAM is fast. The AMD Ryzen platform is similarly well-suited.
(Apple actually tells you how much storage their local model takes up in the settings > general > storage if you're curious)
We can imagine that by 2030 your base model Grandma computer on sale in stores will have at least 32GB of high-bandwidth RAM to handle local AI workflows.
which is why i made the claim that hardware "problem" will be solved in the near future (i don't consider it solved right now, because even the apple hardware is too expensive and insufficient imho), but the more difficult problem of model availability is much, much harder to solve.
There does seem to be a mind share mote, but all you have to do is piss off users a little bit when there's a good competitor. See Digg to Reddit exodus.
Which advertisers would risk having their product advertised by models that have encouraged kids to commit suicide?
This is not mutually exclusive. They have home made robots and let others sell robots on their website. The same way they want to use AI and have resources to make their own. One way to use is to drive those robots. Another to enhance their web site. Current version sucks. I recently return the item because their bot told it has functionality while in fact it didn't.
AWS is very much not the same as Amazon the product selling website.
The two are effectively separate businesses with a completely separate customer base.
> Of course, the AI talent war may end up being an expensive and misguided strategy, stoked by hype and investor over-exuberance.
To me, that's a pretty good explanation.
The world is crazy with AI right now, but when we see how DeepSeek became a major player at a fraction of the cost, and, according to Google researchers, without making theoretical breakthroughs. It looks foolish to be in this race, especially now that we are seeing diminishing returns. Waiting until things settle, learning from others attempts and designing your system not for top performance but for efficiency and profit seems like a sane strategy.
And it is not like Amazon is out of the AI game, they have what really matters: GPUs. This is a gold rush, and as the saying goes, they are more interested in selling pickaxes that finding gold.
I guess Amazon can also probably afford to wait until somebody comes up with an application for AI that is, like, something Amazon can actually sell or use…
Customer service bots? Maybe. Coding bots? I bet they use some internally. Their customers don’t really need them, or if the customer does, the customer can run it on their side.
As I’ve said before, the kind of AI that makes money is called machine learning. Pricing ads, recommending products, improving search, optimizing routing.
In general these fall into the category of things humans cannot do at the scale and speed necessary to run SaaS companies.
Many of the things LLMs attempt to do are things people already do, slowly and relatively accurately. But until hallucinations are rare, slow expensive humans will typically need to be around. The AI booster’s strategy of ignoring/minimizing hallucinations or equivocating with human fallibility doesn’t work for businesses where reliability is important.
Note that ML algorithms are highly imperfect as well. Uber’s prices aren’t optimal. Google search surfaces tons of spam. But they are better than the baseline of no service exists.
You mean something like Kiro?
https://kiro.dev/
AI is huge, it’s just not the only thing happening in tech right now. I say this as an MLE but it seems really unbalanced that LLMs have gotten trillions in investment when other groundbreaking innovations like battery improvements or fusion power or gene therapy have gotten substantially less attention.
Disagree re: DeepSeek theoretical breakthroughs, MLA and GRPO are pretty good and paved the way for others e.g. Kimi K2 uses MLA for a 1T MoE.
Big money investors know that real tangible products that have real tangible benefits aren't usually decimal-point-shifting-your-net-worth jackpots. They make money, sure, but factories can't be built in a day. Also, if they can make AI work as it says on the box, they'll be able to get rid of all those pesky employees and turn their companies into pure money-printing enterprises.
Pay no attention to the cracks that are showing. Nevermind the chill. Everything is fine.
I don't agree, based on my experience trying all Deepseek models on real world software tasks.
AWS specifically have really dropped the ball on this.
I interact regularly with AWS to support our needs in MLOps and to some extent GenAI. 3 of the experts we talked to have all left for competitors in the last year.
re:Invent London this year presented nothing new of note on the GenAI front. The year before was full of promise on Bedrock.
Outside of AWS, I still can’t fathom how they haven’t integrated an AI assistant into Alexa yet either
There's Alexa+ [0] which uses generative AI but it's planned to be a paid option at $20/mo.
[0]: https://www.aboutamazon.com/news/devices/new-alexa-generativ...
> Alexa+ costs $19.99 per month, but all Amazon Prime members will get it for free.
I'm curious if non prime members make up a big market for Alexa. I rarely use my smart devices for anything beyond lights, music, and occasional Q&A, and certainly can't see myself paying 20$/month for it.
I'm curious why anyone would pay $19.99/month for Alexa+ rather than just buy a Prime membership (which is $14.99/month).
Unless of course this is going to be met with a price hike for Prime...
That's what happened with Prime TV, and I absolutely expect it for the AI, too. And it might finally mean I cancel my Prime membership.
Amazon Prime’s price hikes have a predictable cadence: * 2014: $79 to $99
* 2018: $99 to $119
* 2022: $119 to $139
We should expect a price hike from $139 to $159 in 2026, assuming the trend continues.
Meanwhile, Google Fiber has been the same price for 15 years. At least according to the billboard outside my window.
It works out for them because bandwidth gets cheaper over time but inflation eats away at that. $70 today is like $50 back in 2010 when GFiber first launched.
If ChatGPT's Advanced Voice Mode could be served through an always-on device like Alexa, I'd pay for it.
Hmmm... maybe I can install do this through a cheap tablet....
Model sizes have come down enough that it will be possible to run smart home control and simple Q&A entirely locally.
you can do this with home assistant already
Alexa consistently fails with the simplest of questions.
Only thing it can do is set a timer, turn off a light and play music.
It is still nice, but it’s so frustrating when a question pops into my mind, and I accidentally ask Alexa just to get reminded yet again how useless it is for everything but the most basic tasks.
And no, I won’t pay 240 dollars a year so that I can get a proper response to my random questions that I realistically have only about once a week.
> Only thing it can do is set a timer, turn off a light
And it can't even do that without an Internet connection. As someone who experiences annoyingly frequent outages, it never ceases to boggle my mind that I have a $200 computer, with an 8" monitor and everything, that can't even understand "set a timer for 10 minutes" on its own.
Why did you guys bought this in the first place?
Fair question, playing music and timers are nice, but their AI is abysmal so you can imo not ask it to do anything else. I previously worked for a smart home company so I wanted to test out product’s integration with Alexa, so we have some at home. I’m planning to get rid of them, though, and only leave one in the kitchen.
Originally bought it for an elderly parent in assisted living; wasn't as useful as we'd hoped; repurposed it as a kitchen-timer/music-box, for which it works adequately as long as the Internet connection is up. I would not recommend anyone else buy one.
and despite all this, I would pay 240$ a year so that Siri can reliably do what Alexa does today
oh the irony
Alexa has pretty much zero value for me.
Being able to just order something with zero shipping has a ton of value. I could drive down the street but it would still be an hour at the end of the day.
Video streaming has some value but there are a lot of options.
Just pay 0 USD and use Grok app for free?
By far the best thing currently available.
I'm predicting that Grok fails simply due to half (?) the software engineering populating not wanting to use anything Musk has developed.
Grok has to be more than n-times (2x?) as good as anything else on the market to attain any sort of lead. Falling short of that, people will simply choose alternatives out of brand preference.
This might be the first case of a company having difficulty selling its product, even if it's a superior product, due to its leader being disliked. I'm not aware of any other instances of this.
Maybe if Musk switches to selling B2B and to the US government...
If you piss off half of your possible user base, adoption becomes incredibly difficult. This is why tech and business leaders should stay out of politics.
> I'm predicting that Grok fails simply due to half (?) the software engineering populating not wanting to use anything Musk has developed.
I think that's a wildly optimistic figure on your part.
Lets assume that developers are split almost 50/50 on politics.
Of that 50% that follows the politics you approve off, lets err on the side of your argument and assume that 50% of those actually care enough to change their purchases because of it.
Of the 25% we have left, lets once again err on the side of your argument and assume 50% care enough about the politics to disregard any technology superiority in favour of sticking to their political leanings.
Of the 12.5% left, how many do you think are going to say "well, let me get beaten by my competitors because I am taking a stand!", especially when the "beaten" means a comparative drop in income?
After all, after nazi-salute, mecha-hitler, etc blew up, by just how much did the demand for Teslas fall?
The fraction of the population that cares enough about these (on both sides) things are, thankfully, single-digit percentages. Maybe even less.
>>After all, after nazi-salute, mecha-hitler, etc blew up, by just how much did the demand for Teslas fall?
I had been saving up for a Tesla but now I am looking elsewhere. I think a lot of people are doing the same here in Canada. You can grok the actual numbers if you want.
Yeah, a simple example is to just look at how many companies/universities have ChatGPT vs Grok subscriptions internally. I can imagine that many people would have a problem with subscribing to Grok, even if its performance is comparable.
Hmmm, thinking aloud, Oracle?
> This is why tech and business leaders should stay out of politics.
Yeah but they don't stay out of politics, do they? Gemini painting black Nazis was a deliberate choice to troll the vast majority of the population who isn't woke extremists.
My family subscribes to Grok and it's because of politics, not in spite of it. The answer gap isn't large today but I support Musk's goal of building a truth seeking AI, and he is right about a lot of things in politics too. Grok might well fail financially, the current AI market is too competitive and the world probably doesn't need so many LLM companies. But it's good someone wants AI to say what's true and not merely what's popular in its training set.
[flagged]
I think their point was that becoming very involved in politics in a way that alienates half of the population has tarnished Musk’s brand (although, I’d personally adjust that down to more like 1/3). If the point of your whataboutism is that previously it alienated the other 1/3… that doesn’t seem to improve their odds, right?
If anything they’ve now pissed off 2/3 of the population at some point or another.
None of that actually happened.
Literally documented proof.
Literally not. Elon Musk even published the infamous algorithm which he had claimed silenced and censored right wing voices. The only thing the algorithm did that was odd was that it had a special case written into it to boost Elon Musk. You can go look it up.
Mechahitler, the South African genocide debacle, explicitly checking Elons Twitter feed, “You get your news from infowars” system prompts, etc have basically made Grok not a real option for me. I do not want to use a product that is specifically being engineered to be a right wing disinformation machine.
And no, generic brand safety mishaps are not the same; everyone is not doing this.
I enabled Alexa+ few days ago on my devices. Everyone in our home immediately disliked the new Alexa. There were some fairly basic things that Alexa+ cannot do, and Alexa was able to do. Some fairly simple question/answering tasks, and questoins about status of an order.
It will be free if you have a Prime subscription (which means nobody will ever pay for it given Prime is cheaper and you get much more included).
But the project is pretty much dead, it was supposed to launch in February or March and is still not anywhere close to being out.
Yes have seen about that. It’s crazy to me that they still haven’t released it. Really think it could save a dying product
They basically have with Alexa+. It's slightly more limited than ChatGPT, but it sounds much more realistic than stock Alexa and blows it out of the water in terms of smarts. The old model was basically a Siri-like "set timers and check the weather with specific commands," plus some hit-or-miss skills you had to install separately. But the new one gives much more of a sense of understanding your question and can carry on conversations with contextual responses. I've been pretty impressed with it, and the nature of the Echo device makes it much easier to query at will than having to open the ChatGPT app and switch to voice mode.
I agree. I think the Echo devices are good for certain kinds of voice-driven LLM experience. Although it's not that useful for detailed responses and serious questions, since you can't go back and read its response again.
Having briefly interacted with AWS Q out of curiosity, I can see why they haven’t pushed much out publicly. Aside from giving someone a chuckle when they decided to call its suggestions “Q Tips”, it’s functionally useless.
They all but abandoned Astro, their home robot. My suspicion (and information I've heard internally) all but points at them only using Astro as a testbed for self-navigating warehouse robotics, and now that they got what they wanted out of it, the Vesta team basically got thrown to the wolves.
They bought a company that did warehouse robots called kiva before that
https://en.m.wikipedia.org/wiki/Amazon_Robotics
Lingo question: is MLOps like devops for ML, or like flops for ML? I wonder because… actually, either case seems like somewhere Amazon might be losing experts to hot startups.
The former. Basically: build, train, test, deploy, monitor, repeat for ML algos.
As the other response said. It’s DevOps for ML. They have Amazon SageMaker which is the managed ML/MLOps offering that we use extensively because we’re a small team. The documentation is awful
All AWS documentation is awful.
Isn't "Alexa+" doing this? (I have not signed up)
Didnt they basically can most of Alexa a few years ago? I think they realized asking a device questions doesn’t generate profit.
They thought Alexa will enable users to buy more from Amazon just by voice. But most users turned out like me. I would not spend a single dollar on Amazon without actually seeing the item on my mobile or desktop. I wouldn’t even add to cart via Alexa. That’s not an ideal user for device and service that requires hundreds of millions to run.
You saw this with Amazon Dash buttons too. This idea that users would just go "Hey order me some more Tide" and Amazon would just do the right thing at the right price like some sort of intelligent personal assistant. Which it by no means is.
They still have no serverless inference.
SageMaker have serverless inference endpoints
Only if the pipe is defined if you bring your own pipe sage maker offers nothing.
Bedrock is not usable for that?
Not for custom models.
That's because there is no lock-in in the current ecosystems for AI. Yet. But once AIs become your lifetime companion that know everything there is to know about you and the lock-in is maximized (imagine leaving your AI provider will be something like a divorce with you losing half your memory) these parties will flock to it.
The blessing right now is the limit to contextual memory. Once those limits fall away and all of your previous conversations are made part of the context I suspect the game will change considerably, as will the players.
There's a chance this memory problem is not going to be that easy to solve. It's true context lengths have gotten much longer, but all context is not created equal.
There's like a significant loss of model sharpness as context goes over 100K. Sometimes earlier, sometimes later. Even using context windows to their maximum extent today, the models are not always especially nuanced over the long ctx. I compact after 100K tokens.
But you don't have to hold the entire memory in context. You just need to perfect techniques to pull in parts of the context that you need. This can be done via RAG, multi-agent architectures, etc. It's not perfect but it will get better over time.
From my experience context window by itself tells half the story. You load a big document that’s 200k tokens and ask it a question, it will answer just fine. You start a conversation that soon enough balloons past 100k then it starts losing coherence pretty quickly. So I guess batch size plays a more significant role.
By batch size, do you mean the number of tokens in the context window that were generated by the model vs. external tokens?
Because my understandings is that, however you get to 100K, the 100,001st token is generated the same way as far as the model is concerned.
I'm over simplifying here but graph database and knowledge graphs exist. An LLM doesn't need to preserve everything in context, just what it needs for that conversation.
Unless there is a trick that I am missing, I don't think this will work by itself. The fundamental thing is what can the model attend to as it generates the next token.
If you give a summary+graph to the model, it can still only attend to the summary for token 1. If it's going to call a tool for a deeper memory, it still only gets the summary when it makes the decision on what to call.
You get the same problem when asking the model to make changes in even medium-sized code bases. It starts from scratch each time, takes forever to read a bunch of files, and sometimes it reads the right stuff, other times it doesn't.
Context will need to go in layers. Like when you tell someone what you do for a living, your first version will be very broad. But when they ask the right questions, you can dive into details pretty quick.
Export your old chats and put them in a RAG system accessible on the new LLM provider. I did it. I made my chat history into a MCP tool I can use with Claude Desktop or Cursor.
Ever since I started taking care of my LLM logs and memory, I had no issue switching model providers.
Do you have some kind of tooling to automate the process? Would like to try it.
> But once AIs become your lifetime companion that know everything there is to know about you and the lock-in is maximized
Why? It's just a bunch of text. They are forced by law to allow you to export your data - so you just take your life's "novel" and copy paste it into their competition's robot.
It's never quite that straightforward, or perceived as that straightforward. That's why most people just renew their insurance as it's easier than messing about changing and worrying if it will be any better. And how easy is it to transfer emails to another provider?
Who even wants all your previous conversations taken into account for everything you do? How do you grow from never forgetting anything, making mistakes, etc? This is highly dystopian and I sure hope this will forever just be a fantasy.
I have made 100MB of my own chat logs into a RAG memory and was surprised I didn't like using it much. Why? it floods the LLM with so much prior thinking that it loses the creative spark. I now realize the sweet spot is in the middle - don't recall everything, strategic disclosure to get the max out of AI. LLM memory should be like a sexy dress - not too long, not too short. You get the most creative outputs when you hide part of your prior thinking and let the mode infer it back.
I am not an AI enthusiast but I get what you're saying. I occasionally use ChatGPT due to Google being enshittified pretty much. I often do not like the things it tells me and I for sure do not like it complimenting everything I do, but thats something other people seem to like... In my experience starting a fresh chat after a while of back and forth can really help, so I agree with you. Having little to zero prior context is actually the point of view one needs sometimes.
in order for that lifetime companion, we'll need to make a leap in agentic memory.
how do you know memory won't be modular and avoid lock-in?
I can easily see a decentralized solution where the user owns the memory, and AIs need permission to access your data, which can be revoked.
I can easily see a world where users own the devices they buy and install the software they want, but the trend goes in the other direction.
in order for that lifetime companion, we'll need to make a leap in agentic memory.
Well, let’s take your life. Your life is about 3 billion seconds (100 year life). That’s just 3 billion next-tokens. The thing you do on second N is just, as a whole, a next token. If next-token prediction can be scaled up such that we redefine a token from a part of language to an entire discrete event or action, then it won’t be hard for the model to just know what you will think and do … next. Memory in that case is just the next possible recall of a specific memory, or next possible action, and so on. It doesn’t actually need all the memory information, it just needs to know that that you will seek a specific memory next.
Why would it need your entire database of memories if it already knows you will be looking for one exact memory next? The only thing that could explode the computational cost of this is if dynamic inputs fuck with your next token prediction. For example, you must now absolutely think about a Pink Elephant. But even that is constrained in our material world (still bounded physically, as the world can’t transfer that much information through your senses physically).
A human life up to this exact moment is just a series of tokens, believe it or not. We know it for a fact because we’re bounded by time. The thing you just thought was an entire world snapshot that’s no longer here, just like an LLM output. We have not yet trained a model on human lives yet, just knowledge.
We’re not done with the bitter lesson.
Basic questions: what does a GDPR request get you? Wouldn't providers like you to switch to them?
Just look at the smartphone market.
> Once those limits fall away and all of your previous conversations are made part of the context I suspect the game will change considerably, as will the players.
I dunno if this is possible; sounds like an informally specified ad-hoc statement of the halting problem.
Web archive: https://web.archive.org/web/20250901163205/https://www.busin...
Yeah Amazon is massively struggling to hire due to the extremely bad reputation of Andy Jasshole and the RTO 5 policy, and this is not exclusive to AI talents, but is the case for every single role. We have had reqs open for a year in my team and nobody wants to join.
Truthfully, I don't think anyone would recommend their acquaintances to join Amazon right now.
That said, Amazon is actually winning the AI war. They're selling shovels (Bedrock) in the gold rush.
I have had multiple recruiter reachouts from AWS who obviously read my resume and are interested in short cutting me into a senior role at AWS doing interesting things, but at this point AWS reputation is so bad I don’t even entertain such offers.
For senior in-demand talent you are not desperate, and really only desperate people go to work for AWS as they don’t have any better options at a company which respects their employees.
It's not like it had a good reputation earlier either (as a company, perhaps less problematic as an employer). But if I was offered multiple FAANG positions because I had some really attractive skill set, then I'd want a _lot_ more to work at Meta or Amazon than Netflix or Google, just based on my view of the corporate evilness. It's probably completely unfounded, but the fact I have that feeling just shows they haven't taken care of their brand.
I think I’d work at Amazon purely to save the world from the abomination that is cloudformation.
Amazon having trouble to hire I think it is a well deserved result. I hope they never hire great talent again. Lately I heard they're looking for contract hires, which seems to fit their cheapness and lack of ability to attract talent.
At some point, the turnover has to lead to "the blind leading the blind" with nobody having a clearer big picture view on the software they own. This can't be a productive way to run a company, but they seem to persist nonetheless. It may take many years, but I imagine their software will rot from within due to their hiring practices.
Exactly, Amazon practices the equivalent of a decimation of their workforce. This may even work in the initial years, but over time they'll quickly lose their best minds and the software will be unmaintainable.
It's almost funny how they just don't give a shit about being an attractive employer. They never have. Going back to 2002, it's always been "if you don't like it, there's the door."
It seems that they just don't care about the high turnover.
Bedrock is terrible and usage is not high, they cant even serve the anthropic models at scale.
Bedrock? It's like a vibe-coded "router" app. It really doesn't provide anything that is not provided by countless other companies.
AWS is falling behind even in their most traditional area: renting compute capacity.
For example, I can't easily run models that need GPUs without launching classic EC2 instances. Fargate or Lambda _still_ don't support GPUs. Sagemaker Serverless exists but has some weird limits (like 10GB limit on Docker images).
Bedrock is not at all a router. They do provide a routing capability now, but at its core it's a wrapper around models so you can interact with any model with the same unique API.
> For example, I can't easily run models that need GPUs without launching classic EC2 instances.
Yeah okay, but you can run most entreprise-level models via Bedrock.
Only if you want them to go to random inference regions. God forbid you would want inference in a single region. Then you need to be satisfied with 12 month old models that have been superseded 2 times already.
AWS doesn't need to do anything innovative and the enterprises still come. Every product AWS sells has a similar offering from a competitor. But businesses stick with amazon because its all in one. They get bills from one company, trust their security with one company, ect. The only thing that matters to AWS is its reputation.
This works up to a point. I'm extremely familiar with AWS, but we simply _can't_ use it to train our models because it costs 2-3 times more than their competitors. All while requiring us to basically bring up all the infrastructure around maintaining the training cluster ourselves.
Frankly, this is strictly a positive signal to me.
Fargate and lambda are fundamentally very different from EC2/nitro under the hood, with a very different risk profile in terms of security. The reason you can't run GPU workloads on top of fargate and lambda is because exposing physical 3rd-party hardware to untrusted customer code dramatically increases the startup and shutdown costs (ie: validating that the hardware is still functional, healthy, and hasn't been tampered with in any way). That means scrubbing takes a long time and you can't handle capacity surges as easily as you can with paravirtualized traditional compute workloads.
There are a lot of business-minded non-technical people running AWS, some of which are sure to be loudly complaining about this horrible loss of revenue... which simply lets you know that when push comes to shove, the right voices are still winning inside AWS (eg: the voices that put security above everything else, where it belongs).
> Frankly, this is strictly a positive signal to me.
How?
> The reason you can't run GPU workloads on top of fargate and lambda is because exposing physical 3rd-party hardware to untrusted customer code dramatically increases the startup and shutdown costs
This is BS. Both NVidia and AMD offer virtualization extensions. And even without that, they can simply power-cycle the GPUs after switching tenants.
Moreover, Fargate is used for long-running tasks, and it definitely can run on a regular Nitro stack. They absolutely can provide GPUs for them, but it likely requires a lot of internal work across teams to make it happen. So it doesn't happen.
I worked at AWS, in a team responsible for EC2 instance launching. So I know how it all works internally :)
You'd have to build totally separate datacenters with totally different hardware than what they have today. You're not thinking about the complexity introduced by the use of pcie switches. For starters, you don't have enough bandwidth to saturate all gpus concurrently, they're sharing pcie root complex bandwidth, which is a non-starter if you want to define any kind of reasonable SLA. You can't really enforce limits, either. Even if you're able to tolerate that and sell customers on it, the security side is worse. All customer GPU transactions would be traversing a shared switch fabric, which means noisy bursty neighbors, timing side-channels, etc., etc., etc.
> You'd have to build totally separate datacenters with totally different hardware than what they have today.
No? You can reset GPUs with regular PCI-e commands.
> You can't really enforce limits, either. Even if you're able to tolerate that and sell customers on it, the security side is worse
Welp. AWS is already a totally insecure trash, it seems: https://aws.amazon.com/ec2/instance-types/g6e/ Good to know.
Not having GPUs on Fargate/Lambda is, at this point, just a sign of corporate impotence. They can't marshal internal teams to work together, so all they can do is a wrapper/router for AI models that a student can vibe-code in a month.
We're doing AI models for aerial imagery analysis, so we need to train and host very custom code. Right now, we have to use third-parties for that because AWS is way more expensive than the competition (e.g. https://lambda.ai/pricing ), _and_ it's harder to use. And yes, we spoke with the sales reps about private pricing offers.
none of this applies to g6e because it doesn’t have/need a pcie switch, because it doesn’t have rdma support (nor nvlink), which means sriov just works.
And what is your point? What is stopping AWS from using g6e or g6dn on Fargate to keep up with the competitors?
Nothing, but IMO it’s a bad idea. 1. customers who build a compute workload on top of fargate have no future, newer hardware probably won’t ever support it. 2. It’s already ancient hardware from 3 years ago. 3. AWS now has to take responsibility for building an AMI with the latest driver, because the driver must always be newer than whatever toolkit is used inside the container. 4. AWS needs to monitor those instances and write wrappers for things like dgcm.
Fargate is simply a userspace application to manage containers with some ties-in to the AWS control plane for orchestration. It allows users to simply request compute capability from EKS/ECS without caring about autoscaling groups, launch templates, and all the other overhead.
"AWS Lambda for model running" would be another nice service.
The things that competitors already provide.
And this is not a weird nonsense requirement. It's something that a lot of serious AI companies now need. And the AWS is totally dropping the ball.
> AWS now has to take responsibility for building an AMI with the latest driver, because the driver must always be newer than whatever toolkit is used inside the container.
They already do that for Bedrock, Sagemaker, and other AI apps.
> the RTO 5 policy
I'm no expert, but I'm pretty sure this[0] is what RTO 5 is.
[0] https://www.phoenixcontact.com/en-pc/products/bolt-connectio...
RTO 5 is "return to office, 5 days a week"
RTO 996 where it at
[flagged]
Amazon seems to be taking the "When Everybody Is Digging for Gold, It’s Good To Be in the Pick and Shovel Business" approach here.
Don't need to train the models to make money hosting them.
There's also dogfooding though
I am reminded of the Uncomfortable Amazon Truths ( https://news.ycombinator.com/item?id=20980025 ) by Corey Quinn.
While they're protected now, https://news.ycombinator.com/item?id=20980557 quotes the one I recall...
https://threadreaderapp.com/thread/1173367909369802752.html maintains the entire chain of tweets.> Nobody has figured out how to make money from AI/ML
This is clearly not true. Google Ads? Every recommender system? Waymo self-driving? Uber routing algorithms?
If you swapped out ML for LLMs I would largely agree.
I do want to point out that I wrote this in 2019. It was a vastly different landscape, and AWS did a terrible job of promoting the ML value proposition.
I would love to see a retrospective on the AWS uncomfortable truths as a blog post and how they've held out over the past couple years.
2019 was a different time - though I suspect that your statement about making money (as in profit) rather than just revenue (reselling compute for less than you bought it) would hold true for most companies.
That’s an awesome idea. Let me get after it.
One of the debates I have with devops types is if OCI (Oracle Cloud Infrastructure) is a worse name than ICP (IBM Cloud Private).
And would this be admitting defeat to the powers of Terrible Orange Website to get you to write more?
As a side, in 2019 about a week after your tweets I was at a training session for Rancher which worked a reference to one of them into a joke.
I’m always amazed by how far my humor seemed to spread in those days.
Clearly not true anymore given OpenAI and Anthropic's revenue growth.
Revenue... yes. Profit is still an open question.
https://www.cnbc.com/2025/08/08/chatgpt-gpt-5-openai-altman-...
> Last year, OpenAI expected about $5 billion in losses on $3.7 billion in revenue. OpenAI’s annual recurring revenue is now on track to pass $20 billion this year, but the company is still losing money.
> “As long as we’re on this very distinct curve of the model getting better and better, I think the rational thing to do is to just be willing to run the loss for quite a while,” Altman told CNBC’s “Squawk Box” in an interview Friday following the release of GPT-5.
Selling compute for less than it cost you will have as much revenue as you want to pay for.
Anthropic founder described it as: if each model were a company, they be hugely profitable. It looks bad since when the model you trained in 2024 is generating net positive revenue, you’re also training a more expensive model for 2025 that won’t generate revenue until then. So currently, they’re always burning more cash than they’re bringing in, under the expectation that every model will increase revenue even more. Who knows how long that lasts, but it’s working so far.
Paraphrase is from the podcast he was in with the stripe founder, cheeky pints I think
Which is not a good comparison because the LLMs are products not companies. If they are companies, they are competing against each other for revenue.
If I switch from Gemini Pro to Opus, that is good for Anthropic. If I switch from Opus 4 to 4.1, that’s not as good for Anthropic.
Sad that these CEOs can get away with this level of sophistry.
>Revenue... yes. Profit is still an open question.
could have said the same thing about most FAANG companies at one point or another.
The problem for OpenAI and the difference with other FAANGs is that they don’t own the internet. Other companies are able to replicate their product, which prevents them from fully realizing profits.
Google doesn’t have this problem. They only run Google ads in their search results. Same thing for Facebook.
If I have the numbers right, OpenAI will burn more money this year alone than all of those prior companies did in their entire profitless phase of existence.
Their gross profits are very high even though they're not making operating profit.
Isnt amazon basically Anthropic's HW partner very much like OpenAI has microsoft ?
Not only that, I understand they are an investor.
I like how the author threw this in as (nearly) the last sentence:
Of course, the AI talent war may end up being an expensive and misguided strategy, stoked by hype and investor over-exuberance.
AWS has always ridden other products (Postgres, MS-SQL, Redis, etc) that are open source or has negotiated licenses (Windows, MS-SQL, Oracle RDBMS) that are bundled in the end-user price per hour/GB/whatever.
AWS has Bedrock to use various AI providers and has bundled the licensing into the price, so they are getting the users without having to develop the actual AI.
They provide the compute, networking etc, and they provide the users to the AI vendors.
Why would they need to develop their own?
I do not see it as a master strategy. It is just something that happened. Similar to having no plan of having million fakeish Chinese brand selling crap on Amazon.com. But ones they are there then Sure, why not . May be in few year a lot of crashed and burned AI talent will be looking for boring corporate IT/AI job and Amazon will be around to offer that. And if does not happen there will still be ton of other work to do for Amazon.
I don't that's true. Amazon sells infrastructure to other AI companies. If they jump into the model race they become a competitor not an infrastructure provider.
Amazon always competes with everybody. Clients, partners... Everybody.
My intuition is that the root cause it's their frugal culture (frugal as in cheap). They don't want to start a compensation race.
Agree. It would be really messy if they were paying AI engineers multiple time more than regular engineers.
I don't know who needs to hear this but, you can be a big tech company and not compete for every single market the other big tech companies are going for.
I agree. Google killed off a perfectly good product (Google+) just because it could not compete with Facebook.
I and a few others still remember the site fondly, and it had the best UX of any social media service I've used since.
but in practice the firms that missed major technological advances like the internet or mobile did not fare well.
How do you define “missed”? I ask because AI is a bit different than those earlier trends: yes, investors wasted money during the dotcom bubble but there were thousands and thousands of places which spent a modest amount of money and saw almost immediate benefits. AI has had some modest wins but there hasn’t been that “customers love this and are switching to OtherCorp for it” moment - it seems like there’s probably a majority of cases where the executives hyped an AI feature and people yawned and ignored it because it was for the product owner’s CV rather than the users. (Repeat a generation later for the me-too mobile apps which started from an executive’s mandate to “be in the App Store”)
In all of these cases, the problem was losing track of what actually benefits users. AI has that problem really bad now because the infrastructure is expensive and the executive class has been sold on the idea that mass layoffs are just around the corner, and they’re pushing hard to ship before the benefits are there.
> but in practice the firms that missed major technological advances like the internet or mobile did not fare well.
*Microsoft enters the conversation
https://www.zdnet.com/article/microsoft-is-still-paying-for-...
https://www.theverge.com/2023/10/24/23930478/microsoft-ceo-s...
Right, but the firm "Microsoft" is still faring well in spite of missing out on both the major technological leaps at the time.
You have to be joking about them missing the internet?
They became nearly irrelevant because of mobile and had to claw their way back. That is not faring well.
They eventually made it out and survived because of cloud and gaming, but it took what many people consider a major transformation of the company.
Don't let your personal bias about AI cloud the way you see the world.
> You have to be joking about them missing the internet?
No, I'm not. Bill gates famously missed it (and/or severely underestimated the need for internet on Windows PCs) in 1994/5.
Microsoft completely missed the internet, and had to play catchup throughout 1995-1998.
> They became nearly irrelevant because of mobile and had to claw their way back. That is not faring well.
That never happened. They were in no danger at any time. The historic stock price charts, if you care to look them up, would show that the mobile threat you think there was did not even put a blip on their stock price and/or their revenue.
I mean, their revenue never even blipped.
I see what you're saying, but I don't agree with this characterization.
(1) Internet: Netscape came out in 1994, and the internet tidal wave memo was 1995 and internet explorer came out the same year. Windows was rewritten with a focus on the networking stack, with Windows NT coming out in 1993 before the web boom. The internet's value is based on network effects and while you are right that they weren't first to market, they embraced it quickly and if they hadn't it likely would have been disastrous.
(2) Stock price: if you bought MSFT in October the year the iphone came out in 2007, you would take 6 years to break even. If you bought at the top in 2000 you wouldn't break even until 2016. This is a company that was limping along. During the mobile phone boom you'd have been better off putting your money in treasuries than in MSFT.
Yes they survived and were able to do well later. But my original point still stands: if you were running MSFT and wanted to be successful you would have embraced the internet and mobile. Deliberately sitting out a major technological innovation is not a recipe for success because the risk of ruin is very high. And the risk of becoming IBM is even higher.
>(2) Stock price: if you bought MSFT in October the year the iphone came out in 2007, you would take 6 years to break even. If you bought at the top in 2000 you wouldn't break even until 2016. This is a company that was limping along. During the mobile phone boom you'd have been better off putting your money in treasuries than in MSFT.
Using equity returns to claim a business is limping along is bizarre. They were earning $10B profit per year in the early 2000s with 20%+ profit margins, something most businesses can only dream of doing, even today.
https://www.helgilibrary.com/charts/microsoft-corporation-pr...
If that business is limping along, then pretty much all other businesses are on life support.
apple missed the internet amazon missed the mobile
at that given point in time, this was not their main businesses and they fared quite well.
microsoft missing the mobile is different, because mobile being a competitor to desktop destroyed microsoft's main business.
The back-loaded vesting schedule is such blatantly cynical bullshit. It shows that they're planning to overwork you, push you to wash out, and undercompensate you for the experience, which is exactly what I've seen happen to a good number of friends. Amazon has become notorious here in Seattle - everyone knows they're a burnout factory. Some people make it through, and they make good money, but you have to really care about money for that to be worth the effort.
I had an Amazon interview loop on the calendar during my recent job search, a couple of months back, but it was difficult to get excited; they think so very highly of themselves, for what they're offering - and I don't just mean the money, but the culture too. They treat you like an interchangeable wage slave, not like a respected professional; it's all hoops to jump through, and procedures to memorize - dance, monkey, dance!
The recruiter was shocked when I cancelled the rest of the interviews, like, aren't you even going to give us a chance? But no: I had received a good offer from an ambitious, well-organized, well-funded AI startup which was excited to have me on board. With that on the table, why would I put up with Amazon? They won't offer better pay, they can't offer a better culture, and they don't have more interesting problems to work on.
The problem with working at places where you care that much about money is having to work with people who only care about money.
This is a serious challenge in relation to hiring also. If you want to pay for good talent, and so are prepared to pay good money, how do you avoid people who are there for the money.
They got away with this attitude in the earlier days but it’s really hurting them now. A good chunk of the best talent out there won’t even consider Amazon. Culturally it’s very hard to turn that around now and catch up.
90% of the folks there that I know that were good have left for elsewhere. Of the ones that didn’t most are on H1Bs and basically have no choice but to stay and deal with the toxic environment.
This is an uninformed take. Yes the RSU is backloaded. But during the first two years, you get a large monthly cash sign up bonuses so that assuming the stock stays flat, over the four years, your total comp stays flat. If the stock increases your comp goes up.
I spoke to someone who is there now and when you get your yearly review, now you can choose between mostly cash vs mostly stock for your raise and most people choose mostly cash.
I make the same now as I did when I was at AWS and I much prefer my all cash comp over my less cash + RSUs when I was there.
RSU grants assume a growth rate (15%? I forget) so if they stay flat, go down, or grow slower than the baked-in growth rate, then you make less each year. If you do well enough, they’ll give you some RSUs to “make you whole” (as they used to say) but that doesn’t really happen anymore (or not much).
This is not true for your initial four year grant. I’m going to make up a number to make the math easy. Say my total compensation target was $200K. My initial 4 year offer was structured based on the then current stock price.
It would have been what ever it takes where base + prorated signing bonus + RSUs would equal $200K taking into account the 5/15/40/40 RSU schedule.
> The back-loaded vesting schedule is such blatantly cynical bullshit;
I don’t understand the complains about it. Amazon pays monthly cash ”sign-on bonus” in the first two years, which is ~ equal to the stock that you get in the years three and four (counting at the grant price). Is this fact not advertized well enough?
The "sign-on bonus" comes with serious strings attached. A good friend of mine got royally screwed when he mistook that bonus for real money, then got pushed to the point of burnout and had to leave; Amazon demanded a lot of the money back, but he didn't have it anymore.
I worked at Amazon in 2021 and rage-quit after 9 months. The sign-on bonus I received was paid out monthly, so I didn't have to pay anything back. If it's large enough, they pay it monthly because they know it's very likely you won't make it to the 2nd year.
Glad they've fixed that.
(Still, though - why work for people who know they're going to treat you so badly you'll probably have to quit?)
Well for me, I was already 46 when a recruiter from Amazon Retail reached out to me about an SDE (software development) position at Amazon Retail. They said it would require relocation after COVID (this was April 2020). I knew about Amazon’s reputation from both stories and my best friend who had worked as an L6 in the finance department.
There was no way in hell I was going to sell my house and uproot my life to work for Amazon. Then the recruiter after she kept talking suggests I interview for a “permanently remote” [1] “field by design” role at AWS ProServe. I thought sure why not?
The plan was always to make some money - I made over a quarter million more over 3.5 years than I could have made as an enterprise dev working in Atlanta - put AWS on my resume, gain some industry contacts and move on in four years.
I saw the writing on the wall shortly before my 3 year anniversary. I played the game well enough to get past my next vesting period and get my “bust your ass and try to work through your PIP or receive a $40K+ severance and ‘leave immediately’”.
I didn’t hesitate. I took the severance and already had two job offers lined up and had been waiting on the severance offer.
[1] They forced their “field by design” customer facing roles in the office at the end of last year. I would have left anyway before I ever went back into the office.
Amazon doesn't seem to work out for a lot of people. I've tended to have long term jobs and probably wouldn't have been tempted to give them a shot.
Sign-on bonus is prorated and payed monthly, you definitely don’t need to pay back anything (source: I worked at Amazon).
Maybe your friend talked about relocation bonus, which you need to pay back if you don’t work long enough.
My friend is a native-born Seattleite, so no, it was definitely not a relocation bonus.
Perhaps they recently changed their policies? I don't know, but it's not a risk I would want to take. Who would want to work for people who treated their coworkers like that?
Alright, I did some quick research and it seems that they do sometimes pay full first year of sign-on bonus, which you need to repay (prorated). I didn’t see that that during my time at Amazon.
Ah, I just saw this, although I wrote a (much comprehensive) reply here: https://news.ycombinator.com/item?id=45097345
The full payment that requires pro-rates is even worse. They expect you to pay it fully back. (ie. with the deducted taxes included!)
I bet it is possible to profit from a such scheme if Amazon is able to declare that as a reversed-transaction (similar to VAT-refunds) at the end of the fiscal year.
IIRC they pay it out monthly if the bonus is large enough.
I worked there in 2013 and had the signing bonus paid monthly. I thought it was great since I could work there as long as I could tolerate it (10 months) and leave without regrets about having to pay back anything. Decent cash comp so I feel I got a good deal.
I joined in 2022 from a different location, there were 2 kinds of comp in terms of bonues, each split into 2 other;
1. Relocation package a. Lump-sum (7k EUR): You get certain amount of money, and you deal with your own move yourself. (Albeit with some reimbursement possible for the initial trips) b. "Other" (I don't remember the name): More supportive option, good if you have family & furniture to move. They essentially pay everything for you. c. Important: The 7k EUR was subject to the tax, hence I got taxed at 55% (EU) due to having no tax residency at the moment (obviously). Nobody ever mentions this. But the re-payment is with the tax-included, ie. you are expected to pay 7k back! 2. Sign-on bonus: This splits into 2-year period a. 1st year: 50% of the total bonus, transferred to your bank account on your first work day. b. 2nd year: Each month, you get 1/12 of the remaining 50%, essentially something like ~4.18% each month on the second year. c. The 50%/50% ratio may depend on the team/role/location, I heard some of the L4s joined to the team got split of 40%/60% (ie less in the first year) for reasons unbeknownst to me.
Conditions are pretty simple, if you leave (for any reason), you must repay monthly-pro-rated amount that you haven't worked given the total period is 24-months. ie. In Luxembourg, probation is 6-months. (Until) at the end of the probation, Amazon can just fire you for no reason. In this case, since the 2nd year sign-on hasn't vested yet, nothing to pay from that, but you must pay 1/4th of your "relocation expenses" and full half of (ie untaxed full amount divided by 2) sign-on bonus you receive on your first day. (ie. 25% of the total sign-on bonus)
Firstly, I know someone (a Greek national) who left Amazon during his 12th Month. Amazon demanded total of 4k+ euros from the guy, citing he hasn't finished his 12th month, hence the first half of his relocation bonus plus the 1-month of pro-rated sign-on bonus, before tax. As far as I know, it was more or less equivalent to his monthly gross salary, and he paid in installments.
Secondly, I heard someone joined from non-EU country in 2023, and essentially got laid off. But because she was in probation and obviously worker rights are much stricter in EU, Amazon just declared her as a probation-failed case instead of layoff. (She also got laid off within last 2 weeks of her 6-months long probation). Since she only got the residence permit recently, not having more than a few months (when unemployed as a 3rd-country national), plus Amazon demanded money to be paid back. As far as I know she contacted an labour lawyer and they basically advised her to go back and not to pay anything back as it becomes an international matter. And the costs/fees for such is much higher than what would Amazon get it back, hence she did what was suggested. Although it obviously burns the bridges but in this case, Amazon started the fire first...
---
As a result, the practices applied here falls no short of what you can hear from the news. As the company has no heart or soul, people are just numbers in a balance-sheet...
Amazon does not demand your pro rated cash sign on bonus back that you get every pay period for the first two years.
Source: I worked at AWS from 2020-2023.
Glad to hear they fixed this broken policy.
> The back-loaded vesting schedule is such blatantly cynical bullshit.
I don't understand this. A friend was recently offered an insane pay package from Amazon (compared to another big-tech). The way I saw it, the Amazon pay package was more attractive than the alternative because of the back-loaded vesting schedule.
Basically they pay you out in cash for the first two years, then after that you have an option to keep working there. If the stock price goes down in the first two years, you got your guaranteed cash -- no risk (and it would be a good time to interview again). If the stock price goes up, you now have basically an option on extra exposure in the form of staying longer with highly valued RSUs, and now getting some high proportion of your pay in RSUs.
It just seems straight up better? If you want the stock instead of fungible cash, just buy it on the open market?
It's bullshit because it assumes 15% IRR. So if they tell you you're getting $100K in outyear 3, it's not actually $100K, it's $65K of present value equities. If it fails to reach the target value, well, "Ownership" is an LP. You might get some more stock that vests in another year to make up for it, but that assumes you survive the PIP factory for another 12 months.
Oh, and if the stock actually goes up more than 15%, then regardless of your performance you won't get a raise because you've already exceeded band penetration.
Thats not true. They price the stocks at current market value and tell you how many you'll get + what the vesting schedule is.
Amazon is one of the few companies that could benefit the most. Here's an exchange I had with one of their "human" support people:
"Search is broken. If I search for wwvb watch, I get shown tons of watches which are definitely NOT WWVB."
"What browser are you using? Could you try Chrome?"
Sounds like a winning strategy and a money saver to boot.
Exactly! Just build capacity, let other companies duke it out; ultimately they will all likely use AWS for their products anyway.
AWS is more focused on making money off the infrastructure than on the application itself. It took same approach with kubernetes and might I say it has been very successful.
If you are a top AI researcher, there is no good reason to go to Amazon. For what? Pay? Career development? Company prospect? Work-life balance? You get nothing compared to what other companies offer.
And I say, good. We need new, smaller companies with different cultures in this space. We don't want these giant corporations to dominate and control everything.
> We need new, smaller companies with different cultures in this space
we need new, smaller companies with different cultures in every space but won’t be getting any in any space, especially not in this one
AI is full of new and smaller companies. Both OpenAI and Anthropic are quite new, but growing fast.
OpenAI and Anthropic are practically subsidiaries of Microsoft and Amazon. Neither would exist without billions in cloud compute credits from their corporate benefactors. Competing in Generative AI requires the kind of resources that are only available to extremely large and established companies. I do not think all the wrapper companies count when most of them are either being bought out by the big guys or have products that are immediately outmoded in the span of months. Maybe you can make the argument for AI art companies, but Stability basically disintegrated after wasting $100m dollars and Mid journey is directly competing with Google and Meta which is not where I would want to be (aside from running a ghoulishly evil company trying to kill artistic expression).
Sam Altman et al are only “quite new” to my friend’s newborn son :)
new ok.. but smaller? that is not true.
According to Wikipedia, Google has 187,103 employees, Amazon has 1,556,000, and OpenAI has a mere 3,000 employees.
So essentially a lifestyle business - but some people do think they have growth potential.
3000 employees is a lifestyle business? lol. That’s a new one.
In terms of market cap OpenAI (500B valuation) is 5x smaller than Google.
Leaving aside the pendatic "you can't be a multiple smaller than another object", 1/5 the valuation of the 5th most valuable company in the world is probably big enough to qualify you as a big company
> Leaving aside the pendatic "you can't be a multiple smaller than another object"
Feel free to not leave this out, it's a pet peeve of mine. Thank you for the moment of catharsis.
Can you explain this to me? Trying to understand but can’t haha.
Grandparent comment should have said "1/5th the size" instead of 5x smaller.
Oddly we all knew what he meant. Huh.
How small are you? How small are you multiplied by 5?
Market cap doesn't really feel like a good metric of anything other than what it would take to buy a company out. DuPont has a market cap of 30ish billion and 3M around 80B, and both are both larger and frankly more important than probably even Google.
Yeah, the fact that $2.5 trillion of actual investor money chose Google (Alphabet) means very little: what really matters is the opinions of anonymous commenters on HN (especially opinions that start with "doesn't really feel like")
People are so careful when writing anonymous HN comments and so careless in choosing where to invest their own money and the money of funds of which they are the professional manager
> the fact that $2.5 trillion of actual investor money chose Google
Of course, a lot of money invested in Google was invested at a much lower price; if everyone sold all at once you'd have a hard time finding 2.5T of new money to buy all those shares. We could argue about if "not selling" is the same as "choosing again at the new price" every day... but... Google's not the interesting case here anyway.
For a young company in a hot industry like OpenAI total market cap is even less relevant since so much of the company simply isn't liquid anyway and the numbers come from far fewer instances of purchases than for an established public one.
Yes, OpenAI's not being publicly traded makes it harder to value it, but the comment to which I replied referenced 3 publicly-traded companies.
That is a bet and not a metric of company size. Some people bet a lot on small companies, doesn't make them large.
Investors look at how much money is already invested in a company in deciding whether to invest. I.e., investors pay close attention to market cap.
If Google's market cap were $25 trillion, practically nobody would buy Google stock (and practically everyone who already held the stock would immediately sell) because most investors do not believe that Google can ever pay enough dividends or buy back enough stock to justify such a high valuation.
A company's market cap is a collective estimate of how much money the company will to return to investors in the future. When the company is publicly-traded in an open informational regime such as the US, this collective estimate is usually quite "accurate" in the sense that it is very difficult for any single analyst or single team of analysts to improve on the estimate.
An investor can make a big bet on a small company, yes, but the market cap of a company is more than just an indication of how much money has been bet on the company: it also mean that every investor (big or small) who still holds the stock believes that the expected amount of money that company will return to shareholders exceeds the market cap: if there were a holder of Google stock that did not believe that, he would convert the shares into treasury bills or cash in the bank.
Both those companies are 2 orders of magnitude smaller than Amazon or Google.
By what metric? I meant valuation.
OpenAI has 500B valuation, Anthropic has more than 60B.
Private market valuations and public market caps shouldn't be compared. Compare revenue or employee count instead.
Amazon and Apple have never had fundamental research groups. Even before LLMs, the top big-tech fundamental labs were FAIR, Google Research and MSR.
It has never been in Amazon or Apple's DNA to chase a product that doesn't have clear revenue outcomes (as long as adoption lands). AI is no different.
IMO, it's the right decision for Amazon and wrong decision for Apple.
They got burned by over-promising Apple Intelligence and then embarking on a so-far failed, rudderless journey to land features regular people actually gave a shit about. I’m no expert, but I reckon the exact right move is concentrating on their actual deliverable products and features and letting everyone else blow their cash on a maybe months-long moat for dubiously useful advances in categories most of their customers wish everyone would stop talking about.
yeah what IS the deal w/ Apple Intelligence falling off the world?
Beats me. My best guess is they let the hype blind them to the reality that this tech wasn’t merely a few months away from the production-level reliability they needed from it. After a while, it sank in that they couldn’t play this up as a ‘just around the corner’ release and stopped hyping up every useless beta-at-best feature like it was a huge deal. Then, when the “we’re nowhere close” internal communique was leaked, it was officially time to bow out of the hype cycle for a while.
First, I’d like to know what is the deal with Siri not being able to tell me the date without a network connection.
I’m pretty sure it can’t process voice commands on-device other than whichever ones you use to activate it. The thin client architecture was the only option for most of its existence and the only one that works for older devices, which they’ve actually been great about supporting since the battery-life processor throttling bs like a decade ago. It also makes sense that they wouldn’t constantly update their architecture. Now that they’re putting their money on on-device NN stuff I imagine that will change. But let’s be real: using a smartphone without a network connection is an edge case to begin with. On-device processing makes sense for a lot of reasons, but optimizing for smartphone users without data only makes sense if you’ve got a hiking map app or something.
Or providing info from 2016 when asked about US college sizes.
Apple’s biggest problem is their commitment to privacy. Delivering effective AI requires a substantial amount of user data that Apple doesn’t collect.
Their other problem is they value designers and product managers more than engineers (especially top tier AI engineers).
Both problems are basically the death knell of any hope for Apple to have good AI, but combined? It’s never gonna happen. Which is sad because Apple’s on-device hardware is quite good.
Didn’t Amazon invest a total of $8 billion in Anthropic? Seems like a much better choice than trying to do things in-house.
Apple, on the other hand, hasn’t even invested in any of the players.
Was it an investment of actual dollars, or "just" cloud compute credits? Microsoft's investment in OpenAI was over 90% Azure credits, we're told. Which raises the possibility that the whole business is mostly about making your cloud compute business look better than it really is...
Of course it’s probably credits, but it’s still beneficial to both of them — having such huge customers is ideal to stress test your product offering.
Also, just yesterday, they appear to have raised $13B from actual investors, so it seems like they’re going to be fine.
https://www.anthropic.com/news/anthropic-raises-series-f-at-...
That’s an investment that’s going to be a write off.
What about Amazon’s work in formal verification research and Apple’s machine learning research?
[1] https://www.amazon.science/tag/formal-verification
[2] https://machinelearning.apple.com/research
This. It's weird how most of the top tech companies are all morphing into amorphous blobs that want to get into everything and are indistinguishable from each other.
One thought I had recently: Their shareholders are probably mostly the same people. So why even compete?
Aren't every big publicly traded companies shareholders pretty much the same large index fund managers?
Don’t give the FTC any ideas.
actually yeah, let's give them some ideas
Because that's all they know how to do.
Isn’t this something like how everything ends up evolving into a crab?
Carcinisation [https://en.m.wikipedia.org/wiki/Carcinisation], yeah.
Or ‘why every large public company tends to suck the same ways in the US eventually’
Also got to love the linguistic coincidence of Crabs and Cancer and how tech companies grow ever larger (monopolistic) to the detriment of their host (the greater economy/humanity)
It’s not a linguistic coincidence. The disease is named after the animal.
The coincidence is that all animals evolve in a crab-shaped direction (as the meme goes), and all tech companies evolve in a cancer-shaped direction.
That these two “inevitable endpoint things” would happen to be linguistically closely related was unlikely.
It's not a coincidence. Cancer means crab, because the earliest known physicians saw tumors and thought they looked like crabs.
In the US it seems like every company eventually turns into a bank.
My father in law was an IRS Revenue Agent. His quip was that about 20-30% of the civilian economy has tax avoidance as a primary business objective. Real estate is probably the greatest example.
Since financial engineering is in many ways more essential than the actual business. His best example was a chain hotel. In the majority of cases, a typical hotel is a tax vehicle that happens to rent rooms. So no wonder everything becomes a bank. :)
A typical chain hotel (by which I assume you mean a Marriott/Hyatt/Hilton/IHG/Choice/etc brand) is a franchised “small” business.
The franchisee typically pays 10% to 20% royalty to the franchisor (the aforementioned companies). Otherwise, they rent hotel rooms and pay staff to clean them and rent them again.
What is the tax play? That the hotel owner can 1031 into bigger and better hotels? Anyone who owns real estate can do that.
Well, one could argue the entire setup is a means of structuring investments and organizing/attracting Capital eh?
Hotel owner (aka franchisee) puts in capital in a specific way under license, gets help operating it, in exchange for the 10-20% licensing fee paid back to the main corporation.
In many cases, the owner/operator is nearly turnkey, and it’s an effective way of setting up a defacto managed business investment, almost like a LP. Many of the franchised hotels are actually owned/operated by LPs setup for the purpose.
Also in many of these cases, the franchiser provides contacts for financing, may directly facilitate/recruit Capital, and may even provide loans to the franchisee directly.
For most of these larger hotels, the actual act of renting out rooms, etc. is pretty much all automated/managed through the central system anyway, and the majority of the operating costs are structured in such a way as to minimize tax liability.
Is it clearer now?
Not at all. The poster I responded to claimed this:
> a typical hotel is a tax vehicle that happens to rent rooms.
>In many cases, the owner/operator is nearly turnkey,
What does this even mean? Hotels can be turnkey, which in industry terminology means that everything is working sufficiently well such that you can start renting rooms immediately. An owner/operator being turnkey makes no sense.
> setting up a defacto managed business investment
Also makes no sense.
>Also in many of these cases, the franchiser provides contacts for financing, may directly facilitate/recruit Capital, and may even provide loans to the franchisee directly.
Even if true, what does this have to do with taxes?
>For most of these larger hotels, the actual act of renting out rooms, etc. is pretty much all automated/managed through the central system anyway,
No, the actual out of renting out rooms involves housekeepers, maintenance staff, guest service agents, cooks, and management making sure rooms are clean and habitable. Reserving a hotel room is mostly automated, but even that requires a person to manage conflicts of reservations (e.g. unexpectedly needing to extend a stay causing overbooking, changing room types, room locations, etc.)
>and the majority of the operating costs are structured in such a way as to minimize tax liability.
Who doesn't structure their operating costs to minimize their tax liability? If you file married joint instead of married separate or head of household, are you "structuring" your operating costs as a way to minimize tax liability?
The question of how a hotel is used to gain an tax advantage that would otherwise be unavailable remains unanswered.
Most properties are syndicated. Hotels are interesting because they are mix of different asset types. The GP operates the place and LPs contribute capital. Accelerated and bonus depreciation passthrough to the LPs entity.
What does syndicated mean?
And how is a hotel a mix of different asset types?
What does GPs and LPs have anything to do with using a hotel to gain a special tax advantage that is not available to any other commercial real estate?
You should probably do some research. You’re basically asking the equivalent of ‘what is a stock? And why is it different than a bond?’
I am doing research, asking the person who made the claim.
How stocks and bonds come into play is beyond me, unless I am being trolled.
But to summarize, zero evidence of how a hotel is a “tax vehicle”, nor any clarification on what a tax vehicle even is, nor why any other business wouldn’t be able to use the same strategy (if it even exists).
Dude, look up corporate partner structures. General partners. Limited partners. Etc.
Do some basic reading so you can ask informed questions from the answers you have already been given, instead of insisting someone is an idiot when they point out you are not asking useful questions.
And frankly, no one owes you these answers.
Can you give examples? I thought becoming a bank in the US is famously difficult and regulated, so much so that most businesses who can avoid it do so by partnering up with existing, tiny banks. See almost any “fintech” solution, from startups all the way up to Apple.
As far as I understand, becoming a bank is inviting a ton of overhead with little profit potential.
I don’t think they meant a literal bank, but finance games become a bigger part of their core strategy. For example, AirBnB for a while made a majority of its profits by investing the money guests paid during the gap between booking and actual stay (paying the host).
Correct - at some point, the enterprise revolves around either finding better return on excess Capital they have, or finding additional Capital.
Which is the core premise of a bank, even if the business doesn’t say ‘Bank’ on the side of the building.
Agreed but elsewhere in threat people are assuming a literal bank with all the regulations that go along with it, etc.
True, though…. [https://en.m.wikipedia.org/wiki/GE_Capital]
Tesla has huge crypto holdings.
This is kind of true I never thought about it until now
Nit: Trees/Grass are even more of this than Crabs.
The two strategies for plants are to grow super tall to absorb the sun, or super wide (and small) to.... absorb the sun.
Tall needs wood or other 'strong' polymer to support height. Short and wide is perhaps weak from an individual level but far more efficient.
And trees and grass respectively have such genetic diversity that it's clear that none of these damn plants are of the same genetic line.
Nit: grasses are a distinct genetic lineage, the Poaceae family. There are a few other linages outside of Poaceae that have convergently evolved to look like grasses, sedges and rushes, but they all fall in the same clade, Monocots.
Trees, on the other hand, are a growth habit, exhibited by species in a wide variety of plant families, even grasses (e.g palm trees).
There's no point in discussing a meme, but carcinisation doesn't occur in that wide of a range, and of course the reverse phenomenon (decarcinisation) is also observed.
It's a fun image, but just as Facebook isn't becoming Apple, and Amazon won't become OpenAI, evolution phenomenons are more complex than "everything becomes X"
They’re just turning into conglomerates.
It was common in the post wwii era in America and its Asian allies like Korea with its chaebols and Japan with its somethings I can’t remember the name of. The Asian countries forms were normally based around a single family, we’ll need more time with the current US form to see if they are also dynastic
The Japanese, family-owned, generally pre-WWII conglomerates were called zaibatsus. After WWII they were (nominally) dissolved and the now more loosely connected groups of companies are called keiretsus.
Stakeholders expect (and price assets for) endless growth.
This is the unfortunate answer to a lot about why companies do
We are all addicted to growth - everyone is chasing the hockey stick curve which means a business that provides a stable business and grows modestly is seen as a failure in some parts
I don't know, there doesn't seem to be much overlap to me. Apple is a hardware business, Microsoft is software, Google is search, Facebook is social media, Amazon is distribution and compute. They do have their fingers in each other's pies but not to a large extent.
Which is a godsend for the users. Can you imagine a world where there is only one big cloud provide, say AWS, and all the big companies with the infra just sit out? Can you imagine how expensive AWS would be and how much power it has over the users?
They'll just buy the competition once it seems like it's at a good price. Capitalism leads to concentration.
Power accretes.
That’s what happens when you print trillions of dollars. Suddenly investors have too much Monopoly money and they want to spend it on something, anything, that might not make as much of a loss as holding cash during the subsequent inflation.
Yeah, they will come, new companies from China, that will eat the market too, with their beautiful 996 work life balance, and we will go back to growing corn.
As a bonus you will have a very long vacation.
We, the tech, are literally a leftover of the once overwhelming engineering superiority of the west that will shrink in the next 5 years.
The problem is that when you start that smaller company and it gets successful, you will be acquired. Big companies rarely build things anymore.
Just FTR - it's VERY rare for people to come up with more than one winning idea
Once a company gets big off its grand idea, there's little to no chance of it having another big winner, so buying one is best (and its cheaper too, you know it's a good idea, and you don't have to spend so much R&D on it.
Big companies have never built anything new. It goes back decades. before tech companies, it was giants like GE who grew through acquisition after acquisition and eventually imploded from the incompetence blob (which takes a long time to accumulate the damage). The same will happen to the current big tech companies in a few decades.
Sure they do. Amazon built AWS well after it was big. Apple built the iPhone. Microsoft built VS Code. just to name a few examples.
It makes some sense to sell out if you're building a product that will at best acquire a tiny sliver of the market, which almost all companies will. But there's at least a few AI companies, like Anthropic, that could potentially balloon towards becoming a Big Tech company. So it makes sense for them to not sell out for the time being.
Sell out, or get big enough to buy out others.
You don'thave to - but that means setteling for a job that earns an okay income. Sell out for millions now - more that your lifetime earnings and use the time and money for - what you want
I wonder what percentage of people in tech are working to cash out some way. 60%? More?
I argue it is both understandable (autonomy is a healthy thing) and also damages the culture at large.
Most by far are working for someone else. They get no stock option, or if they get them they are of minimal value. Their generally get a good 401k (us only) and so can retire well off but would not call themselves rich.
>you will be acquired.
You say this as if it's a coercive given, when you could just as easily say.. Nope, and continue to see how you compete with some agility. It might fail, but most of the big tech companies currently acquiring smaller companies themselves started small with acquisition offers being rejected along the way. Sure, there's selection bias at work there, but there are also many cases of smaller to mid-size companies that also said no to acquisition and still managed to find their successful niche.
Being acquired is not a given and neither is failure if you do compete in some way with the megacorps.
I see nothing about the current tech landscape that at all distinguishes it from previous landscapes in which smaller companies succeeded AND rejected acquisition.
> You say this as if it's a coercive given, when you could just as easily say.. Nope, and continue to see how you compete with some agility.
It’s the same framing as calling offering someone a higher salary as “poaching” like we’re property being stolen by one lord from another.
Looking at you Steve Jobs and your anti poaching agreement
I wish more people said no. However, the reality is it appears to be a given. If you're offered millions and millions of dollars, most people do not say no. The world is worse for it, but it's the truth.
I think you really need to reconsider the definition of a given. Because most people don't say no to a thing doesn't make it coercive. Whether the world is worse or better for it in this context, I can't be sure (though I lean more towards big corporations eating nearly all competitors as generally dangerous), but we're still talking about voluntary choices in a market where competition still does frequently emerge to overhaul what's established. It's impressive how often people don't notice this even as it happens all around them in ways that directly benefit their daily lives.
They sponsor your visa. That’s it.
[dead]
Smaller companies will still be VC backed, which limits the variation you'll see in culture
If that is how you feel, then the reason for why it currently is the way it is should not give you much comfort. It's not like Amazon can not decide to change things and throw more money at the issue from a different angle in the future.
The current SoA AI requires massive investment and CPU time (which isn't free)
No matter who is funding that, they are going to be pushing hard for a return (ell, unless they like money going up in smoke)
Yep Amazon should be split up. No reason that AWS, Alexa, satellite internet, their online store, and groceries have to be one company.
But is there grounds to say that as a conglomerate they pose a large harm to market health to merit a breakup? For example, few regulators want to break up Mondragon.
Even if you’re a mid-tier AI researcher… or even a hobbyist one, I can’t see a good reason to go to Amazon.
AWS has now become one of the most hated tools, right next to Jenkins.
Amazon is turning into a dinosaur like Cisco or IBM.
There's no value in Amazon burning money to 'compete' when there no clear endgame. Right now the competition seems to be who can burn a a hundred billion dollars the fastest.
Once a use case and platform has stabilized, they'll provide it via AWS, at which poiny the SME market will eat it up.
Not only that, but all the compute spent, and hardware bought, will be worthless in 5 years.
Just the training. Training off of the internet! Filled with extremists, made up nuttery, biased bs, dogma, a large portion of the internet is stupids talking to stupids.
Just look at all the gibberish scientific papers!
If you want a hallucination prone dataset, just train on the Internet.
Over the next few years, we'll see training on encyclopedias and other data sources from pre-Internet. And we'll see it done on increasingly cheaper hardware.
This tiny branch of computer sciences is decades old, and hasn't even taken off yet. There's plenty of chance for new players.
How exactly do you foresee "pre-internet" data sources being the future of AI.
We already train on these encyclopedias, we've trained models on massive percentages of entire published book content.
None of this will be helpful either, it will be outdated and won't have modern findings, understandings. Nor will it help me diagnose a Windows Server 2019 and a DHCP issue or similar.
We're certainly not going to get accurate data via the internet, that's for sure.
Just taking a look at python. How often does the AI know it's python 2.7 vs 3? You may think all the headers say /usr/bin/python3, but they don't. And code snippets don't.
How many coders have read something, then realised it wasn't applicable to their version of the language? My point is, we need to train with certainty, not with random gibberish off the net. We need curated data, to a degree, and even SO isn't curated enough.
And of course, that's even with good data, just not categorized enough.
So one way is to create realms of trust. Some data trusted more deeply, others less so. And we need more categorization of data, and yes, that reduces model complexity and therefore some capabilities.
But we keep aiming for that complexity, without caring about where the data comes from.
And this is where I think smaller companies will come in. The big boys are focusing in brute force. We need subtle.
New languages will emerge or at least versions of existing languages till come with codenames. What about Thunder python or uber python for the next release.
I like AWS overall.
(Though I'm pretty familiar with some of the concepts, I know some things to avoid (e.g., "push this button to set up a very expensive global enterprise scale observability platform of numerous complicated services, because you asked about a very simple turn-key syslog service"), and I'm expecting the occasional configuration headache (and, lately, configuration wizard bugs).)
For a new startup, I'd use AWS for all serving and hosting purposes by default, iff you have someone who can avoid pitfalls, and handle problems.
If you don't have such a technical person, maybe start off with managed Kubernetes service with high-level UI, at AWS or one of the other cloud providers, and try not to make too big a mess (which might slow you down, or take you down) before you can afford to hire specialists to make sure it keeps working for you.
I still like AWS all these years later. It’s trusted in the enterprise and you can empower people to do what they need to themselves with IAM. And it’s pretty reliable.
> AWS has now become one of the most hated tools
By whom? Certainly no one I work with. AWS has some sharp edges and frustrations but we couldn't do half of what we do without it.
https://www.reddit.com/r/csMajors/s/WPI0L2Pzsj
This is a weird take. I don’t know any developers who hate AWS. It’s the dominant cloud provider for a reason.
Since when? It’s extremely popular
It’s not, it’s just become something forced on orgs like confluence
I don't think cloud computing is a hated "tool", it's effectively taken over running on-prem.
It's the same as saying buying electricity from a network is worse than having your own generators.
You say Jenkins is hated, but surely it's no more hated or worse than any other bigger player in the space like Teamcity or Bamboo.
huh how did Jenkins all of a sudden get into this discussion? And why the hate, it was king of CI for over a decade and for good reason.
> AWS has now become one of the most hated tools
News to me
[dead]
Amazon would have a lot of trouble competing for top talent due to their reputation.
No one is looking at this issue correctly. Saying out of the AI "talent war" is a smart move. AI is due to collapse under its own weight.
1) High-quality training data is effectively exhausted. The next 10× scale model would need 10× more tokens than exist.
2) The Chinchilla rule. Hardware gets 2× cheaper every 18 mo, but model budgets rise 4× in that span. Every flagship LLM therefore costs 2× more than the last, while knock-off models appear years later for pennies. Benchmark gains shrink and regulation piles on. Net result: each new dollar on the next big LLM now buys far less payoff. The "wait-and-copy" option is getting cheaper every day.
I usually do not agree with the Amazon leadership (well, recently they haven't been "Right A lot"!)
But I agree with the following statement Matt Garman gave recently;
It's because AI usually creates slop, without review these "slop" build up. We don't have infinite context window to solve the slop anyway. (even if we do, the context-rot has been confirmed)Also, on average, Indian non-Tech employees who manages thousands of spreadsheets or manually manages your in-store cameras are much more cheaper than the "tokens" and the NVIDIA GPUs you can throw at the problem, at least for now and a foreseeable future.
> It's because AI usually creates slop
I don't think his point was we should hire junior engineers because they're cheap and lean into AI and AI produces slop. His position is not that he wants to cheaply create slop.
He wants to hire people who are cheap and love using AI because he sees that as a better long term strategy than making senior engineers embrace AI late into their career.
I agree but it doesn’t seem to be intentional on their part.
from outside looking in i think this is a move i would support if i were in amazon leadership. let the other players pay for the AI movement, pick up the fruits of their labor a couple of years down the line. i dont think amazon's main play is AI anyways, if anything it's to facilitate AI with their complementary platforms in AWS
Why would they? I hear their main revenue is in AWS and AdTech. Assuming this is true, why would they need bleeding edge AI?
Amazon also didn't read the room when it fired most of its Alexa staff just as GenAI was taking off.
https://www.cnbc.com/2023/11/17/amazon-cuts-several-hundred-...
Of course not being able to monetise Alexa has always been a problem, but these and the article's issues are all to do with poor planning and top tier business direction.
AWS dropped the ball but they didn’t really try. Apple OTOH…
Apple finance took over and scoffed at their engineers wanting to spend $10B on AI hardware, so they bought back stocks instead because Buffet understands that.
MS, Google, and Meta are spending hundreds of billions on AI, however their stocks are growing by the same amount. This doesn't seem like a crazy spending when looking at this, and better than just buying stock back and doing exactly nothing like Apple is doing.
Apple can afford to take their time, people will keep buying phones no matter whose AI they run on them. Google on the other hand can see its search business evaporating within a few years.
Of course they've stood it out. The rate of change and the R&D expenditure is off the charts. It buys them marginal utility to hire AI talent at incredible salaries to keep them at table stakes.
Meanwhile, the models are getting larger and more complex, with more users, putting the support infrastructure well beyond what individuals and even small companies can afford to outright buy. You can easily spend well over a million on even basic infrastructure to try to support some of the newer models and make it available to a few end users.
As a point of strategy for individuals and small entities, it really is cheaper in this case to spin up some AWS instances for a bit to do some LLM work and then spin them down when not in use.
So if you were AWS do you mine for gold? Or do you sell shovels?
That whole “sell shovels” thing never really made sense, even in the pre-GPU hyperscaler days. BTW, the shovel is GPU (owned by NVidia for now).
AWS, Azure, GCP weren’t just renting servers. They built whole platforms - databases, ML stacks, dev tools, security. Way more than shovels.
The moat was owning the stack. MS used Azure to power Office and now Copilot. Google used infra to juice Search, YouTube, Ads. Even Amazon used it for retail + Alexa. They were mining gold and selling shovels.
And raw compute was never where the money was. Renting VMs was the cheap layer. The profits came from all the higher level services built on top.
Now with AI it’s even more obvious:
Models drive the workloads. OpenAI/Anthropic/DeepMind aren’t just customers, they’re shaping the infra itself. Whoever owns the models sets the rules.
No models = no moat. If AWS isn’t building frontier models, it’s just reselling Nvidia GPUs while MS + Google wrap their clouds around first party models + SDKs. That pulls customers deeper into their stacks, not Amazon’s.
Falling behind compounds. Training/deploying models forces infra breakthroughs (chips, compilers, scaling). If AWS isn’t in that game, they’ll eventually struggle to even run other ppl’s models as well as rivals.
So if Amazon “sits this one out,” it’s not just losing bragging rights. It’s giving up control of the future of compute.
I agree with you that the profit comes from higher level services built on top.
But I think you are making it sound like Amazon's moat is that it came up with its own technology behind its services.
A lot of times AWS was just grabbing a bunch of popular open source stuff off the shelf and hosting it (e.g., RDS, EKS, etc). Yes there is some R&D work but almost none of what Amazon has come up with is rooted in their own work.
The value they give you is the hosting, maintenance, and compliance of all these services. If you're paying AWS extra to host your database on RDS or your Kubernetes cluster in EKS, you're generally not paying AWS to come up with a better database than anyone else, you're just paying them to help you manage permissions, backups, replication, and other maintenance/compliance/management issues that a company needs for its internal services.
In other words, Amazon's AI customers don't need Amazon to build models. They just need Amazon to use someone else's models, host them on private enterprise compute that easily ties in to existing infrastructure, RBAC, etc, and make everything compliant and easy to maintain. A whole lot of the value is being able to answer audits with "AWS handles our database backups/data security/etc" rather than saying "we have a great ops team and here's all our proof that we handle our database backups/data security/etc properly."
I think it's actually explicitly Amazon's job to sit this one out, especially since they never successfully made a good business or consumer ecosystem device like a smartphone or PC operating system.
Are modals the future of compute?
I’m not 100% convinced this is true. Additionally, I’m not convinced that a waiting pattern right now sets Amazon up for a point of no return. It seems plausible for Amazon to pull an Apple here, to wait until technology is more mature and use their unique position to provide a quality offering.
The problem is Amazon is usually Apples non competitive cloud partner. They can’t both sit this out. AWS needs to learn in a hurry whether they should be in the model business to supply Apple with Siri LLMs. Bc if not Apple is going to Google (and Google cloud). Thats not good for AWS. Amazon is in a bit of a bind bc they should be acquiring Anthropic but not at bubble prices.
I don't understand why AWS needs to be in the model business. They didn't develop databases, they didn't develop Kubernetes, they didn't develop Linux, and the list goes on and on.
Not a whole lot in their portfolio actually has a lot of Amazon technology behind it. They've got some mild forks here and there, and they've got some stuff like Fargate that has AWS R&D work behind it but piggybacks concepts/tech stacks that definitely didn't originate from Amazon.
A lot of their value has really nothing to do with developing the underlying technology.
Is Apple really going to shovel a bunch of money to a direct competitor (Android) in a way that is likely to result in less differentiation for Apple mobile devices?
Android isn’t really a competitor it’s an ecosystem. Google Pixel is a competitor. And would Apple start funding Google’s Pixel market penetration in exchange for AI? Perhaps. Google was funding Apple with default search money for a long time. I think the 20B payment flow is about to reverse direction, not disappear.
This follows my take, in terms of where the profitability will be long term. It will be with the hardware vendors, and not the model creators. Time will see if I'm right, but, as hard as it is to create a good model, it does seem to be something that can be replicated by others.
Not exactly going out on a limb here by predicting hardware manufacturers are going to make money on AI
True, but I'm arguing that it won't be nearly as profitable to be on the software end.
More cynically, the way to get rich during a gold rush is to sell shovels.
If AWS instances in this analogy are shovels, what are the GPUs?
GPUs are the land that's staked. They're of limited intrinsic value in the sense that they're in the way of getting to the gold (models), but there's a small supply of it so they go for a lot on the market. But if there was literally any other way to make the models even for a few percent cheaper, they model builders would move to that. On the inference side, most of the cloud providers are looking for pretty much any way to server that up more cheaply, with custom TPUs, or other tensor units of some type.
We saw this with crypto mining where truckloads of expensive GPUs were dumped in the trash after the proof of work became so hard it became not worth the cost of electricity to keep on that generation of card.
One layer up, the blade. Silicon would be the iron in this analogy.
The physical server itself would be the wooden handle, I guess.
Picks or any of the other required equipment.
Raw iron, timber, & coal supplies.
it's also a "moneyball" situation - hire hyper expensive AI stars or get 10-fold? 100-fold? cheaper smart AI folk sans star power
> Blah blah blah, money this, money that
See, that’s the problem with what Amazon has done to you. It’s always about money with you guys. Good research is about the opposite of money. The people who don’t know what that means, who can’t fathom to understand what “the opposite of money” means without turning everything into a contrived story about money: they can’t do good R&D. Every single great R&D director will tell you this, and a bunch of people will downvote this comment, who have never been in a meaningful R&D role.
A good research culture is capable of listening to broad, generalized, completely accurate criticism in public and not downvote. Downvoting is your problem guys!
OpenAI has a million little haters out there and do you know how much time their people spend downvoting comments online? Zero. And honestly they’re paid way better than the poor souls who have wound up at Amazon, so it’s really, truly the case that none of this money money money culture really adds up to much for the little guy.
If there’s any one person to point the finger at - like why does Amazon, with its vast resources and tremendous talent, produce basically zero meaningful publicly influential research - it’s Jeff Bezos. You’re talking about strategy? The guy in charge is a colossal piece of shit, with a piece of shit girlfriend and a piece of shit world view, at least as bad as Larry Ellison, whose only redeeming factor is that MacKenzie Scott is a much smarter person than he ever was.
They invested $8B in Anthropic so they will be OK.
That’s because LLMs are just snake oil… Look at what a flop ChatGPT 5 was. If someone manages to actually make something useful, Amazon has a stake in Anthropic. Otherwise why should they waste their money while competitors bankrupt themselves at scale over a hype cycle.
> But actually, Amazon, Apple etc aren't natural homes for this, they don't need to burn money to chase it.
Why wouldn't consumer AI be a natural home for Apple?
Apple is constantly under blast for being slow to AI but if you look at the current state of AI, it feels like something Apple would never release -- the quality just isn't there. I don't necessarily think Apple only dipping their toes into AI is that poor of a decision right now. They still have the ability to blow the roof off the market with agents and device integration whenever the tech is far enough along to be trustworthy to the average consumer.
Apple's natural home is hardware, and the consume software integrated with that hardware. Off the top of my head I can't think of any "hard" software created by Apple, it's all about the UX and the integrations.
So unless Apple thinks it can outcompete it's BigTech competitors in something it historically hasn't done much of, best leave it to them.
> Off the top of my head I can't think of any "hard" software created by Apple, it's all about the UX and the integrations.
This sounds like you’re either unfamiliar with what software they make or underestimate the complexity of things like a modern operating system. For example, most people would consider Swift hard, or the various Core frameworks, or things like designing a new modern file system and doing in place migrations on billion devices, etc.
creative talents needs more comfortable and unconstrained working environment.But that against and broke many Amazon‘s leadership principal. Also amazon dont want to pay more money to those expensive ai talents.
AI directly threatens and can enhance revenue for companies like Google and Meta, so it makes sense for them to invest in those areas. It's relevant for user segmentation, ad targeting, content creation/user engagement, and even search. LLMs are fantastically powerful search and discovery tools.
Amazon though, sells physical goods and access to physical servers. Whatever is going on with AI, Amazon will profit from without having to burn its own money in advancing SOTA.
> Amazon doesn't provide useful tools for building durable multi-AZ applications. Most customers are not going to implement Paxos
Don't really agree here, yes they screw you financially on cross-AZ bandwidth, but all of their popular services are built to work well across availability zones.
Most people don't need access to a low level consensus service like Paxos, instead they will be using one of amazons managed database services, or s3, that provides higher level abstractions, and automatically manages consensus behind the scenes.
Unless it’s documentDB in which case it manages it front of the scene by just freezing the entire db for minutes on end
As far as I understand it, Anthropic is effectively “Amazon AI” similar to how OpenAI is “Microsoft AI”. So they are not sitting it out, they’re very involved.
Well Amazon is mostly guided by pragmatism rather then than hype, so there they are, waiting for the dust to settle to see what directly helps their bottom line.
Amazon still has huge R&D spends, always had. Bezos had a dictum around having a high experimentation (and failure) rate as a matter of principle. They may not be making news-making moves, but I'm sure they'll develop the muscle in AI. Probably - just really honing on the customer use cases and working backwards over the long term.
It is interesting but I think they're doing the right thing. AWS bedrock works pretty well and you can access frontier models plus everything open source on it. In the end, I was disbelieving that Graviton would be good but the latest r8g series are great for compute so I imagine GPU compute will similarly be mastered by them in time.
Amazon is in the business of providing picks and shovels for LLMs.
Why should they need to develop their own models?
AI will subvert and destroy Amazon's internal management culture, where status is gatekept by who can write the best 6-page reports to read before the meeting!
Or more likely -- Amazon management knows just how hard writing actually is, how hard to produce something with clarity and signal instead of just common-knowledge cliches, and so they understand that this LLM wave is overhyped. They're letting the other big players do the hard work, and effectively selling LLMs short by abstaining from the race.
I think Amazon and Apple see who is doing the "work" in commerce and manufacturing and they know and realize that some non deterministic AI is not a big threat. Sure it creates nice text, video or image but that is not "work" for these small company eating giants. They know that work counts with real goods moving in the real world, energy moving and robots that can actually act with certainity (99,999% time like internet, web as a tech ?)
Interesting theory, but Amazon uses tons of stochastic methods (including deep reinforcement learning) throughout the business, including warehouse inventory management. "Determinism" is not some north star that operations people always adhere to, because the physical world is deeply stochastic and pretending it isn't does not make for a successful operations career.
There’s Gaussian and fractal randomness. Fraud and transportation losses are Gaussian, for example - they average out to known values. An empowered LLM can wreck absolute havoc, and if it’s not empowered there’s no reason to spend $100b on training it.
This really isn’t highlighted enough. Most real world probabilities that are evaluated follow a Gaussian structure. LLMs…don’t? Fractal probably? Heavy tailed maybe (like a Cauchy distribution)? But certainly not in ways that companies are currently accustomed to.
My favorite science fiction threat is an AI able to hallucinate an OS so well any hardware rectangle could be used.
Somebody needs to update that sci-fi story “BLIT”
I think Apple is just sitting idly and waiting for AI to mature to "just works" level without all the potential legal and PR minefields. It's too wild and unpredictable of an experience for their buttoned down, bland and inclusive corporate image. Apple may soon find itself producing very capable but dumb bricks if they don't catch up. Google can and will go all out on AI in Android at some point.
Amazon I think just hasn't understood how to cohesively integrate AI into their offerings. Meanwhile they're selling shovels to the prospectors with AWS.
I guess both of these understand the Ai moat is not very large, and don't buy into AGI dreams.
Assuming most people will want more AI in Android, doesn't seem so popular shoehorned into Windows 11.
It has much better utility in a phone (accessibility to camera and photos, various sensors, contacts, chats, smart home, payment methods...) than on a PC. I can imagine an AI that's more proactive, I don't go to ask a question but it helps me manage my day effectively and get more information where its useful.
Okay, but does it need to be deeply integrated into the OS or can it just interact with programs through their normal interfaces?
The most effective way to get an LLM to control a computer right now is to just give it a unix terminal because it's already a text-based environment where programs are expected to be highly interoperable.
What I'm saying is that you don't need to stop everything to redesign around AI, just allow for a decent level of interoperability that iOS (and largely android) doesn't currently have.
The mobile app development model is oriented around packaging somewhat useful software (that could usually be a web app) with malware and selling it for $0.99, necessitating a ton of sandboxing and preventing this type of interoperability in the first place. I would say focus on the semantic HTML aspect of the web and design some way for LLMs to interact with websites in an open-ended way.
A reasonable theory. Apple does hardware and supply chains, and sees how far there is to go. Nvidia does hardware too, but it's profiting hugely from the AI boom and has no reason to push back.
How do you explain the Elon keiretsu, though? Tesla and SpaceX are pretty tethered to the physical world, and in theory should have visibility into the same discrepancies that Apple sees. So why is Elon pushing so hard to develop Grok? Is it just ideology for him, or what?
Who knows what’s going on in that mad man’s mind!
Elon, like everyone, is smart at some things and dumb at others. When you realize that about the world, it will help you learn from the smart sides of folks.
what's Elon smart at though? so i can learn..
Despite his mad and destructive social and political side, as an engineer and business man he is extremely smart and effective.
He makes lots of unnecessary major and cringy mistakes in both engineering and business too, but his net on both counts is astounding.
And while he may overuse it for PR, he has put himself at great financial risk when pushing through major capability developments and business hurdles. His rewards were earned.
But the sick picture of the richest person in the world, spamming stupidity, and harming countless numbers of people's lives in order to prop up his juvenile ego is hard to look past for many. For good reason.
He is a strong mix of both extremes of capability/impact spectrum, not just one.
Reality distortion. Financial shenanigans. Hiring people who can execute.
And, despite all the haters, he does understand rocket science pretty well, and rocket economics even better.
Eh, not really. He managed to hire people who can manage him well enough to get him out of the way.
Spin it any way you like, but he's still hiring people who deliver.
It's not really spin. He's the wallet, not the talent.
So why did Bezos get nowhere with Blue Origin despite throwing more money at it? Or every car manufacturer that tried to build EVs before Tesla? Or every satellite internet provider before Starlink?
Apparently nearly zero people can do this.
(2012) https://www.northwestern.edu/magazine/spring2012/feature/roc...
> Shotwell had lunch with a co-worker who had just joined the then-startup company SpaceX. They walked by the cubicle of CEO Elon Musk. “I said, ‘Oh, Elon, nice to meet you. You really need a new business developer,’” Shotwell recalls. “It just popped out. I was bad. It was very rude.” Or just bold enough to capture Musk’s attention. He called her later that day in 2002 and recruited her to be vice president of business development, his seventh employee.
Can you imagine something like that working today?
I've heard he's pretty smart at making money.
The recent Tesla figures indicate otherwise.
Politics. Not the kind of politics that makes people like you, but the kind of politics that gives you power.
IMO Grok is the downstream consumer result of internal investment in AI at Twitter. One of the first things Elon did after buying was put all the useful APIs behind a paywall, which would be a reasonable first step if you bought it in part for the enormous training data the platform generates every day and wanted to limit competitors' access to it. Grok is then mostly just a way to get feedback on the tech.
Twitter and Tesla have two interesting datasets for AI.
[dead]
[dead]
Hey! Your words echo this IMO! If not, strongly recommended :) https://www.notboring.co/p/the-electric-slide
Yeah they have a rigid structure of document writing, most of which is now obsolete.
I’d argue it’s the opposite. Good documents are far more important when everyone can generate garbage.
You have all these labs spending billions on researchers and training clusters, not seeing much return on investment, meanwhile Amazon just partners with the labs, and provides inference for their models, and that seems to be fairly profitable for them.
Why buy the cow if you can get the milk for free?
Amazon will reap the results of the talent war. Why?
Because Amazon will build services on top of the technologies that come out.
Just today on hn there was a guy that trained his tiny model and got better results than most of the big models. He wasn’t paid 200m.
The gold rush is here, but the results are still shaking out.
I had totally forgotten that I signed up for the Kiro waitlist. It seems Amazon has also totally sat out the interest in their AI offering.
Has anyone had a chance to use Kiro at all? At this point I'm not even interested in it anymore, even if I got an invite.
Why would Amazon want to invest in AI, which would help their customers find what they want to buy, when they make all their profit from showing their customers what they don’t want to buy with monetized search?
The top researchers published enough details on how to build what works well. Amazon can copy what's useful. They'll probably do it in a way that makes profit, too. Neither talent wars nor AI, startup models contribute to that.
Amazon just has to host LLaMa and Qwen locally, just as they do so many other packages developed by others, and charge for their AWS compute credits. Why do they need “AI talent”?
> Amazon is exploring ways to allow for more "location-flexible" roles, the document added.
If only the technology existed to do work remotely, what a shame.
if it existed, amazon doesn't have it... (rip chime)
AWS is a fantastic metaphor for selling shovels during a goldrush. I can see how Amazon would not be motivated to go out and "look for gold" itself.
They are in the shovel sales business in this gold rush.
[dead]
Zuckerberg has to rush into every new fad with billions of dollars because he was a one trick pony with FaceBook.
Metaverse will never be FaceBook.
I am sure that Amazon is making big piles of money proving hardware to run AI compute and related infrastructure products.
Is there any reason to think that you can't just show up late and do your thing after everyone clears a path?
Amazon doesn't need people with particular credentials on the org chart to bring in VC money. They already have plenty.
“Never interrupt your opponent while he is in the middle of making a mistake” — Shang Tsung
I think Amazon is definitely adopting the philosophy of selling shovels to gold miners.
The Nova line of models, the AWS LLMs, are horrible. So yeah, figures.
It's probably trying to figure out what it can sell at a large margin.
Isn't amazon the metaphorical guy selling shovels at the gold rush.
> The company has flagged its unique pay structure, lagging AI reputation, and rigid return-to-office rules as major hurdles.
No mention of reputation for harsh/ruthless/backstabby management practices towards employees (including for tech white collar, not just biz and blue collar)?
Is that not a major factor? Or are they not aware of it? Or is mentioning it politically off-limits? Or is putting it in writing a big PR risk? Or is putting it in writing a big legal risk?
I know Amazon's reputation for treating employees poorly came up in multiple discussions at one university's big-name AI lab, for example. Not only do some people read the news, but people talk, in groups and privately.
After reading so many horror stories (whether actually true or not), my mind now just associates working at Amazon with mostly negatives: They're going to ride you like a horse and beat you up, for below-average compensation, and then if you want to claw your way up, it's a Game Of Thrones style slugfest with few winners. The opposite of "Rest and Vest." If this is exaggerated, they sure aren't doing any PR work to deny it or counter this negative reputation.
They also produce legions of managers who get fed up working for amazon and leave for greener pastures which they then turn toxic.
I have personally seen this happen (ex-Amazon managers coming in and turning the place toxic) at 2 separate companies now
I’ve heard multiple recruiters (from different agencies in different geos) refer to them as “amholes” and said they’re hard to place and difficult to break their bad habits.
I've bumped into a lot of execs who say they don't want to hire ICs or managers (usually only one or the other) coming from specific big-name companies, and will instruct external and internal recruiters/HR and hiring managers about that.
Not big-name companies in general, but specific companies among them.
It seems to be about belief of culture taint risk (e.g., the way engineering is done, or the misaligned careerism or sharp-elbowedness that's promoted by the company). Though there's also sometimes a belief that particular large companies hire lots of people who aren't good (only, apparently, at LeetCode interviews).
I'm a bit sympathetic to those theories, though I personally don't rule out any individual. I think, say, all the FAANGs do also have individual people who are capable and well-intentioned, and haven't been permanently branded with whatever problematic culture of the company they're at.
(Though there was a time when I thought a person wouldn't have gone to one particular social media company unless they were either a sociopath or completely unaware of news in the real world, but it's more nuanced now. And there's currently an aggressively pro-fascism company that AFAICT never should've seemed like a good idea to anyone who wasn't evil or oblivious, though, I have to remember that they like to hire "impressionable children", and we now have tech track undergrads who haven't had time for anything but STEM classes and LeetCode since early teens, so they might be forgiven. I was recently considering denylisting anyone who'd gone to a different tech company, which had a well-known decades-long history of chronic underhandedness, but then I saw that a colleague who'd majorly helped me out once had finally gone there. Which is another lesson to myself not to generalize in ways unfair to the individual.)
If you want to hire people who share your values and your values include moral responsibility for the megacorp one works for, you’re right not to hire from companies you feel are immoral.
I personally don’t ascribe corporate amorality (as opposed to immorality) to all who work for it and thus with narrow exceptions would blacklist someone for working at a company who, e.g., has a CEO I dislike, practices wage suppression, etc.
I would strongly consider against hiring someone who worked in certain addiction based industries such as tobacco or gaming (not gamedev, the other kind).
I think it's unfair to apply these rules to culture-takers like ICs but an exec who has been there a while? Their CV should go straight in the trash.
Hiring ex-Amazon managers credibly signals to capital markets that a tech company past its hype phase is getting serious about controlling costs and disciplining lazy or entitled engineers. It’s in their interest to have this reputation.
> they sure aren't doing any PR work to deny it or counter this negative reputation.
They don't seem to give a shit. In the retail space their name means "low quality Chinese counterfeit products with fake reviews" and I've seen no effort on Amazon's part to counter that perception either.
> for below-average compensation
Maybe compared to FAANG, but not compared to most corporate developer jobs out there.
It’s better for them if they have this reputation. It lets them select for desperate people.
At least in Europe I have a lot of trouble finding companies paying me as much as Amzn does, and it’s not even close.
Perhaps working for American companies remotely will change that view, but it’s too much a hassle for me at the moment.
amazon is known to be lower end of FAANG, its more a logistic company tbh
logistics in terms of hardware and software not necessary bleeding tech in giants club
That’s far from the truth..
so you didn't read the original article????
So you are basing your comment on what Amzn does based on comments from HR?
Please…
AWS does a lot of bleeding edge stuff, many of which never make it to prod.
"AWS does a lot of bleeding edge stuff"
apparently this bleeding edge tech is basically a low tech in another FAANG company
sorry, they are not in the same league
but if its for producing AWS slop service, amazon win. I can give you that
Sorry that’s just hilarious.
> unique pay structure
As an ex-Amazonian, I hate seeing this corporate euphemism. We would be reminded yearly that compensation at Amazon was “peculiar”, when really it was just relatively low for FAANG. I would have preferred frank honesty, which I think would look like “we pay relatively low wages, for relatively good engineers, and the difference makes more money”
That used to be the case but as of a couple years ago (maybe 2023?) pay packages got bumped and in terms of TC Amazon is very competitive now. You'll likely get a better offer than Google in cash value. But the non-TC benefits are really really bad (no free food, 5 day RTO, oncall policies, etc). For those reasons I think most would take a Meta or AI lab offer over Amazon right now if they're willing to grind.
Interesting, one would think that would mean easier interviews and whatnot so as to allow for greater number of applicants and churn, but it is not what I have heard about it.
it is a good place for new graduates to solve some challenging engineering problems at scale and learn. Most of the employees do not last more than 2 years. People who stick for longer, admire that type of culture and are made for amazon. Their stock has also performed extremely well.
I had senior faculty in a top-5 CS program tell me that they steered their students away from AWS because they didn't want them to be miserable.
That's not a bug, but a feature
The also wisely sat out of the Crypto and NFT hype cycles
would have thought that Amazon's API-first internal systems would make it a wonderful case for AI. Unfortunately doesn't seem to be the case.
Under Andy Jassy’s watch Amazon really just missed the boat on AI big time. Alexa was a huge missed opportunity. They had a huge foothold and then basically did nothing with it. AWS doesn’t really have compelling AI offerings. Bedrock should be good but is a mess. The GPU offerings struggle to keep pace with competitors.
The rambling answer to the “why are you behind” question on the last earnings call indicates it’s a sore spot for leadership, but at this point it’s too little too late. The best talent has already settled elsewhere. The only real saving grace is that if/when the AI bubble pops being so far behind might not be a terrible thing.
Didn't Amazon aquihire Adept Labs?
There's a special place in Metaverse for people who don't jump on the latest hype train.
Eh, looks at my aws bedrock bill, I think they are doing alright.
Amazon, and AWS specially, just don’t have recognized leaders in this space at the helm. I think that’s OK as they should focus on the more boring but important infrastructure stuff.
Jassy’s long rambling answer on the last earnings call though does suggest that being way behind on AI is a sore spot for leadership.
Attracting top talent though is a challenge for Amazon beyond just AI. Their reputation has become a real issue and the top folks simply have better options.
My head hurts reading a lot of this AI-like drivel. All I see is a bunch of options for better search engines here. That said, with all this crap feeding off itself, it's not going to get any better.
My bet is on Apple's upcoming announcement.
Um, Amazon has invested $8B in Anthropic.
I think they learned some hard lessons from Alexa.
Meh. Amazon's strategy is to hire grad students and PhD's from UoW to come and work for them for two years before they go back to academia.
So what's the tldr, are they just too cheap too pay for top AI scientist talent-- which is imagine they would need in order to enter the fray?
Apple: Hold my beer!
[dead]
[dead]
> "GenAI hiring faces challenges like location
No! Really? With RTO? Unbelievable /s
[flagged]
I've been having great success with bottles (https://usebottles.com/). The fact that you can have multiple tweakable prefix helps a lot.
I find AWS extremely difficult to use compared to GCP. Even though we received startup credits—which are essentially free money—we’re letting them go to waste because the platform is so much harder to work with.
It’s no surprise that AWS’s revenue growth is lagging behind GCP and Azure.
Beyond the AI talent gap, Amazon seems to be making serious missteps in its own core business.
It reminds me of Apple. At first, people thought Apple was being strategic by staying out of the AI race and waiting to pick the winner. But in reality, it turned out to be an inability to adapt to the new trend. I expect the same pattern from Amazon.