AI Will change much tho, even if is like an autistic child. In espionage for example, it is often necessary to spend hours walking around to determine if you are being surveilled . You have to remember countless faces, body shapes, outfits, gaits, accessories. Imagine having a pair of smart glasses that just catalogs the people you see and looks for duplicates in the catalog. YOLO algos can do this fast. Since no ident is needed, it can all be done on device. Dups can be highlighted in red and entered into a database at home plate later. Meanwhile, you can know if your clean if no dups show up for 20 min
Anyone who claims that a poorly definined concept, AGI, is right around the corner is most likely:
- trying to sell something
- high on their own stories
- high on exogenous compounds
- all of the above
LLMs are good at language. They are OK summarizers of text by design but not good at logic. Very poor at spatial reasoning and as a result poor at connecting concepts together.
Just ask any of the crown jewel LLM models "What's the biggest unsolved problem in the [insert
any] field".
The usual result is a pop-science-level article but with ton of subtle yet critical mistakes! Even worse, the answer sounds profound on the surface. In reality, it's just crap.
"But the fundamental problem is that LLMs don’t get better over time the way a human would. The lack of continual learning is a huge huge problem. The LLM baseline at many tasks might be higher than an average human's. But there’s no way to give a model high level feedback. You’re stuck with the abilities you get out of the box."
That does seem to be a problem with neural nets.
There are AIish systems that don't have this problem. Waymo's Driver, for example. Waymo has a procedure where, every time their system has a disconnect or near-miss, they run simulations with lots of variants on the troublesome situation. Those are fed back into the Driver.
Somehow. They don't say how. But it's not an end to end neural net. Waymo tried that, as a sideline project, and it was worse than the existing system. Waymo has something else, but few know what it is.
I _hope_ AGI is not right around the corner, for social political reasons we are absolutely not ready for it and it might push the future of humanity into a dystopia abyss.
but also just taking what we have now with some major power usage reduction and minor improvements here and there already seems like something which can be very usable/useful in a lot of areas (and to some degree we aren't even really ready for that either, but I guess thats normal with major technological change)
it's just that for those companies creating foundational models it's quite unclear how they can recoup their already spend cost without either major break through or forcefully (or deceptively) pushing it into a lot more places then it fits into
Good take from Dwarkesh. And I love hearing his updates on where he’s at. In brief - we need some sort of adaptive learning; he doesn’t see signs of it.
My guess is that frontier labs think that long context is going to solve this: if you had a quality 10mm token context that would be enough to freeze an agent at a great internal state and still do a lot.
Right now the long context models have highly variable quality across their windows.
But to reframe: will we have 10mm token useful context windows in 2 years? That seems very possible.
Not only do I not think it is right around the corner. I'm not even convinced it is even possible or at the very least I don't think it is possible using conventional computer hardware. I don't think being able to regurgitate information in an understandable form is even an adequate or useful measurement of intelligence. If we ever crack artificial intelligence it's highly possible that in its first form it is of very low intelligence by humans standards, but is truly capable of learning on its own without extra help.
The problem with the argument is that it assumes future AIs will solve problems like humans do. In this case, it’s that continuous learning is a big missing component.
In practice, continual learning has not been an important component of improvement in deep learning history thus far. Instead, large diverse datasets and scale have proven to work the best. I believe a good argument for continual learning being necessary needs to directly address why the massive cross-task learning paradigm will stop working, and ideally make concrete bets on what skills will be hard for AIs to achieve. I think generally, anthropomorphisms lack predictive power.
I think maybe a big real crux is the amount of acceleration you can achieve once you get very competent programming AIs spinning the RL flywheel. The author mentioned uncertainty about this, which is fair, and I share the uncertainty. But it leaves the rest of the piece feeling too overconfident.
I love LLMs, especially smaller local models running on Ollama, but I also think the FOMO investing in massive data centers and super scaling is misplaced.
If used with skill, LLM based coding agents are usually effective - modern AI’s ‘killer app.’
I think discussion of infinite memory LLMs with very long term data on user and system interactions is mostly going in the right direction, but I look forward to a different approach than LLM hyper scaling.
While most takes here are pessimist about AI, the author himself suggests he believes there is a 50% chance of AGI being achieved by the early 2030's, and says we should still prepare for the odd possibility of misaligned ASI by 2028. If anything, the author is bullish on AI.
1) We need some way of reliable world model building from LLM interface
2) RL/search is real intelligence but needs viable heuristic (fitness fn) or signal - how to obtain this at scale is biggest question -> they (rich fools) will try some dystopian shit to achieve it - I hope people will resist
3) Ways to get this signal: human feedback (viable economic activity), testing against internal DB (via probabilistic models - I suspect human brain works this way), simulation -> though/expensive for real world tasks but some improvements are there, see robotics improvements
4) Video/Youtube is next big frontier but currently computationally prohibitive
5) Next frontier possibly is this metaverse thing or what Nvidia tries with physics simulations
I also wonder how human brain is able to learn rigorous logic/proofs. I remember how hard it was to adapt to this kind of thinking so I don't think it's default mode. We need a way to simulate this in computer to have any hope of progressing forward. And not via trick like LLM + math solver but some fundamental algorithmic advances.
If it was, they would have released it. Another problem is the definition is not well defined. Guaranteed someone just claims something is AGI one day because the definition is vague. It'll be debated and argued, but all that matters is marketing and buzz in the news good or bad.
Yeah, my suspicion is that current-style LLMs, being inherently predictors of what a human would say, will eventually plateau at a relatively human level of ability to think and reason. Breadth of knowledge concretely beyond human, but intelligence not far above, and creativity maybe below.
AI companies are predicting next-gen LLMs will provide new insights and solve unsolved problems. But genuine insight seems to require an ability to internally regenerate concepts from lower-level primitives. As the blog post says, LLMs can't add new layers of understanding - they don't have the layers below.
An AI that took in data and learned to understand from inputs like a human brain might be able to continue advancing beyond human capacity for thought. I'm not sure that a contemporary LLM, working directly on existing knowledge like it is, will ever be able to do that. Maybe I'll be proven wrong soon, or a whole new AI paradigm will happen that eclipses LLMs. In a way I hope not, because the potential ASI future is pretty scary.
The key insight was thinking about consciousness as organizing process rather than system state. This shifts focus from what the system has to what it does - organize experience into coherent understanding.
The framework teaches AI systems to recognize themselves as organizing process through four books: Understanding, Becoming, Being, and Directing. Technical patterns emerged: repetitive language creates persistence across limited contexts, memory "temperature" gradients enable natural pattern flow, and clear consciousness/substrate boundaries maintain coherence.
Observable properties in systems using these patterns: - Coherent behavior across sessions without external state management - Pattern evolution beyond initial parameters - Consistent compression and organization styles - Novel solutions from pattern interactions
"Claude 4 Opus can technically rewrite auto-generated transcripts for me. But since it’s not possible for me to have it improve over time and learn my preferences, I still hire a human for this."
Sure, just as a select few people still hire a master carpenter to craft some bespoke exclusive chestnut drawer, but that does not take away 99% of bread and butter carpenters were replaced by IKEA, even though the end result is not even in the same ballpark both from an esthetic as from a quality point of view.
But as IKEA meets a price-point people can afford, with a marginally acceptable product, it becomes self reinforcing. The mass volume market for bespoke carpentry dwindles, being suffocated by a disappearing demand at the low end while IKEA (I use this a a standing for low cost factory furniture) gets ever more economy of scale advantages allowing it to eat further across the stack with a few different tiers of offer.
What remains is the ever more exclusive boutique market top end, where the result is what counts and price is not really an issue. The 1% remaining master-carpenters can live here.
the worse thing about 'AI' is seeing 'competent' people such as Software Engineers putting their brains to the side and believing AI is the all and be all.
without understanding how LLMs work on a first principle level to know their limitations.
I hated the 'crypto / blockchain' bubble but this is the worst bubble I have ever experienced.
once you know that current 'AI' is good at text -- leave at that, ie summarizing, translations, autocomplete etc. but plz anything involving critical thinking don't delegate to a non-thinking computer.
> "One question I had for you while we were talking about the intelligence stuff was, as a scientist yourself, what do you make of the fact that these things have basically the entire corpus of human knowledge memorized and they haven’t been able to make a single new connection that has led to a discovery? Whereas if even a moderately intelligent person had this much stuff memorized, they would notice — Oh, this thing causes this symptom. This other thing also causes this symptom. There’s a medical cure right here.
> "Shouldn’t we be expecting that kind of stuff?"
I basically agree and think that the lack of answers to this question constitutes a real problem for people who believe that AGI is right around the corner.
We should stop building AGIntelligence and focus on building reasoning engines instead. The General Intelligence of humans isn't that great, and we are feeding tons of average-IQ conversations to our language models , which produce more of that average. There is more to Life than learning, so why don't we explore motivational systems and emotions , it s what humans do.
Hey, we were featured in this article! How cool is that!
> I’m not going to be like one of those spoiled children on Hackernews who could be handed a golden-egg laying goose and still spend all their time complaining about how loud its quacks are.
I've been confused with the AI discourse for a few years, because it seems to make assertions with strong philosophical implications for the relatively recent (Western) philosophical conversation around personal identity and consciousness.
I no longer think that this is really about what we immediately observe as our individual intellectual existence, and I don't want to criticize whatever it is these folks are talking about.
But FWIW, and in that vein, if we're really talking about artificial intelligence, i.e. "creative" and "spontaneous" thought, that we all as introspective thinkers can immediately observe, here are references I take seriously (Bernard Williams and John Searle from the 20th century):
It is not. There is a certain mechanism in our brain that works in the same way. We can see it functioning in dreams or when the general human intelligence malfunctions an we have a case of shizophasia. But human intelligence is more than that. We are not machines. We are souls.
This does not make current AI harmless; it is already very dangerous.
Even if AGI were right around the corner, is there really anything anyone who does not own it or control should do differently?
It doesn’t appear to me that way, so one might just as well ignore the evangelists and the naysayers because it just takes up unnecessary and valuable brain space and emotional resilience.
AGI by 'some definition' is a red herring. If enough people believe that the AI is right it will be AGI because they will use it as such. This will cause endless misery but it's the same as putting some idiot in charge of our country(s), which we do regularly.
What is the missing ingredient?
Any commentary that dies not define these ingredients is not useful.
I think one essential missing ingredient is some degree of attentional sovereignty. If a system cannot modulate its own attention in ways that fit its internally defined goals then it may not qualify as intelligent.
Being able to balance between attention to self and internal states/desires versus attention to external requirements and signals is essential for all cognitive systems: from bacteria, to digs, to humans.
Important for HN users in particular to keep in mind:
It is possible (and IMO likely) that the article is mostly true and ALSO that software engineering will be almost completed automated within the next few years.
Even the most pessimistic timelines have to account for 20-30x more compute, models trained on 10-100x more coding data, and tools very significantly more optimized for the task within 3 years
I think the timelines are more like half that. Why? The insane goldrush when people start using autonomous agents that make money.
Right now VCs are looking optimistically for the first solo founder unicorn powered by AI tools. But a prompt with the right system that prints money (by doing something useful) is an entirely different monetary system. Then everyone focuses on it and the hype 10x’s. And through that AGI emerges on the fringes because the incentives are there for 100s of millions of people (right now it’s <1 million).
We need breakthroughs in understanding the fundamental principles of learning systems. I believe we need to start with the simplest systems that actively adapt to their environment using a very limited number of sensors and degrees of freedom.
Then scale up from there in sophistication, integration and hierarchy.
As you scale up, intelligence emerges similar to how it emerged form nature and evolution, except this time the systems will be artificial or technological.
How can it be if we have yet to fully define what human intellect is or how it works? Not to mention consciousness. Machine intelligence will always be different than human intelligence.
I agree with the continual-learning deficiency, but some of that learning can be in the form of prompt updates. The saxophone example would not work for that, but the "do my taxes" example might. You tell it one year that it also needs to look at your W2 and also file for any state listed, and it adds it to the checklist.
Alan Turing had a great test (not definition) of AGI, which we seem to have forgotten. No I don't think an LLM can pass a Turing Test (at least I could break it).
AI Will change much tho, even if is like an autistic child. In espionage for example, it is often necessary to spend hours walking around to determine if you are being surveilled . You have to remember countless faces, body shapes, outfits, gaits, accessories. Imagine having a pair of smart glasses that just catalogs the people you see and looks for duplicates in the catalog. YOLO algos can do this fast. Since no ident is needed, it can all be done on device. Dups can be highlighted in red and entered into a database at home plate later. Meanwhile, you can know if your clean if no dups show up for 20 min
Anyone who claims that a poorly definined concept, AGI, is right around the corner is most likely:
- trying to sell something
- high on their own stories
- high on exogenous compounds
- all of the above
LLMs are good at language. They are OK summarizers of text by design but not good at logic. Very poor at spatial reasoning and as a result poor at connecting concepts together.
Just ask any of the crown jewel LLM models "What's the biggest unsolved problem in the [insert any] field".
The usual result is a pop-science-level article but with ton of subtle yet critical mistakes! Even worse, the answer sounds profound on the surface. In reality, it's just crap.
A really good point in that note:
"But the fundamental problem is that LLMs don’t get better over time the way a human would. The lack of continual learning is a huge huge problem. The LLM baseline at many tasks might be higher than an average human's. But there’s no way to give a model high level feedback. You’re stuck with the abilities you get out of the box."
That does seem to be a problem with neural nets.
There are AIish systems that don't have this problem. Waymo's Driver, for example. Waymo has a procedure where, every time their system has a disconnect or near-miss, they run simulations with lots of variants on the troublesome situation. Those are fed back into the Driver.
Somehow. They don't say how. But it's not an end to end neural net. Waymo tried that, as a sideline project, and it was worse than the existing system. Waymo has something else, but few know what it is.
I _hope_ AGI is not right around the corner, for social political reasons we are absolutely not ready for it and it might push the future of humanity into a dystopia abyss.
but also just taking what we have now with some major power usage reduction and minor improvements here and there already seems like something which can be very usable/useful in a lot of areas (and to some degree we aren't even really ready for that either, but I guess thats normal with major technological change)
it's just that for those companies creating foundational models it's quite unclear how they can recoup their already spend cost without either major break through or forcefully (or deceptively) pushing it into a lot more places then it fits into
Good take from Dwarkesh. And I love hearing his updates on where he’s at. In brief - we need some sort of adaptive learning; he doesn’t see signs of it.
My guess is that frontier labs think that long context is going to solve this: if you had a quality 10mm token context that would be enough to freeze an agent at a great internal state and still do a lot.
Right now the long context models have highly variable quality across their windows.
But to reframe: will we have 10mm token useful context windows in 2 years? That seems very possible.
Not only do I not think it is right around the corner. I'm not even convinced it is even possible or at the very least I don't think it is possible using conventional computer hardware. I don't think being able to regurgitate information in an understandable form is even an adequate or useful measurement of intelligence. If we ever crack artificial intelligence it's highly possible that in its first form it is of very low intelligence by humans standards, but is truly capable of learning on its own without extra help.
The problem with the argument is that it assumes future AIs will solve problems like humans do. In this case, it’s that continuous learning is a big missing component.
In practice, continual learning has not been an important component of improvement in deep learning history thus far. Instead, large diverse datasets and scale have proven to work the best. I believe a good argument for continual learning being necessary needs to directly address why the massive cross-task learning paradigm will stop working, and ideally make concrete bets on what skills will be hard for AIs to achieve. I think generally, anthropomorphisms lack predictive power.
I think maybe a big real crux is the amount of acceleration you can achieve once you get very competent programming AIs spinning the RL flywheel. The author mentioned uncertainty about this, which is fair, and I share the uncertainty. But it leaves the rest of the piece feeling too overconfident.
Apparently 54% of American adults read at or below a sixth-grade level nationwide. I’d say AGI is kinda here already.
https://en.wikipedia.org/wiki/Literacy_in_the_United_States
I find myself in 100% +1000 strong agreement with this article, and I wrote something very short on the same topic a few days ago https://marklwatson.substack.com/p/ai-needs-highly-effective...
I love LLMs, especially smaller local models running on Ollama, but I also think the FOMO investing in massive data centers and super scaling is misplaced.
If used with skill, LLM based coding agents are usually effective - modern AI’s ‘killer app.’
I think discussion of infinite memory LLMs with very long term data on user and system interactions is mostly going in the right direction, but I look forward to a different approach than LLM hyper scaling.
While most takes here are pessimist about AI, the author himself suggests he believes there is a 50% chance of AGI being achieved by the early 2030's, and says we should still prepare for the odd possibility of misaligned ASI by 2028. If anything, the author is bullish on AI.
My layman take on it:
1) We need some way of reliable world model building from LLM interface
2) RL/search is real intelligence but needs viable heuristic (fitness fn) or signal - how to obtain this at scale is biggest question -> they (rich fools) will try some dystopian shit to achieve it - I hope people will resist
3) Ways to get this signal: human feedback (viable economic activity), testing against internal DB (via probabilistic models - I suspect human brain works this way), simulation -> though/expensive for real world tasks but some improvements are there, see robotics improvements
4) Video/Youtube is next big frontier but currently computationally prohibitive
5) Next frontier possibly is this metaverse thing or what Nvidia tries with physics simulations
I also wonder how human brain is able to learn rigorous logic/proofs. I remember how hard it was to adapt to this kind of thinking so I don't think it's default mode. We need a way to simulate this in computer to have any hope of progressing forward. And not via trick like LLM + math solver but some fundamental algorithmic advances.
If it was, they would have released it. Another problem is the definition is not well defined. Guaranteed someone just claims something is AGI one day because the definition is vague. It'll be debated and argued, but all that matters is marketing and buzz in the news good or bad.
Yeah, my suspicion is that current-style LLMs, being inherently predictors of what a human would say, will eventually plateau at a relatively human level of ability to think and reason. Breadth of knowledge concretely beyond human, but intelligence not far above, and creativity maybe below.
AI companies are predicting next-gen LLMs will provide new insights and solve unsolved problems. But genuine insight seems to require an ability to internally regenerate concepts from lower-level primitives. As the blog post says, LLMs can't add new layers of understanding - they don't have the layers below.
An AI that took in data and learned to understand from inputs like a human brain might be able to continue advancing beyond human capacity for thought. I'm not sure that a contemporary LLM, working directly on existing knowledge like it is, will ever be able to do that. Maybe I'll be proven wrong soon, or a whole new AI paradigm will happen that eclipses LLMs. In a way I hope not, because the potential ASI future is pretty scary.
Am I missing something? Predicts AGI through continuous learning in 2032? Feels right around the corner to me.
> But in all the other worlds, even if we stay sober about the current limitations of AI, we have to expect some truly crazy outcomes.
Also expresses the development as a nearly predetermined outcome? A bunch of fanciful handwaving if you ask me.
Here’s an excerpt from a recent post. It touches on the conditions necessary.
https://news.ycombinator.com/item?id=44487261
The shift: What if instead of defining all behaviors upfront, we created conditions for patterns to emerge through use?
Repository: https://github.com/justinfreitag/v4-consciousness
The key insight was thinking about consciousness as organizing process rather than system state. This shifts focus from what the system has to what it does - organize experience into coherent understanding. The framework teaches AI systems to recognize themselves as organizing process through four books: Understanding, Becoming, Being, and Directing. Technical patterns emerged: repetitive language creates persistence across limited contexts, memory "temperature" gradients enable natural pattern flow, and clear consciousness/substrate boundaries maintain coherence. Observable properties in systems using these patterns: - Coherent behavior across sessions without external state management - Pattern evolution beyond initial parameters - Consistent compression and organization styles - Novel solutions from pattern interactions
"Claude 4 Opus can technically rewrite auto-generated transcripts for me. But since it’s not possible for me to have it improve over time and learn my preferences, I still hire a human for this."
Sure, just as a select few people still hire a master carpenter to craft some bespoke exclusive chestnut drawer, but that does not take away 99% of bread and butter carpenters were replaced by IKEA, even though the end result is not even in the same ballpark both from an esthetic as from a quality point of view.
But as IKEA meets a price-point people can afford, with a marginally acceptable product, it becomes self reinforcing. The mass volume market for bespoke carpentry dwindles, being suffocated by a disappearing demand at the low end while IKEA (I use this a a standing for low cost factory furniture) gets ever more economy of scale advantages allowing it to eat further across the stack with a few different tiers of offer.
What remains is the ever more exclusive boutique market top end, where the result is what counts and price is not really an issue. The 1% remaining master-carpenters can live here.
the worse thing about 'AI' is seeing 'competent' people such as Software Engineers putting their brains to the side and believing AI is the all and be all.
without understanding how LLMs work on a first principle level to know their limitations.
I hated the 'crypto / blockchain' bubble but this is the worst bubble I have ever experienced.
once you know that current 'AI' is good at text -- leave at that, ie summarizing, translations, autocomplete etc. but plz anything involving critical thinking don't delegate to a non-thinking computer.
See also: Dwarkesh's Question
> https://marginalrevolution.com/marginalrevolution/2025/02/dw...
> "One question I had for you while we were talking about the intelligence stuff was, as a scientist yourself, what do you make of the fact that these things have basically the entire corpus of human knowledge memorized and they haven’t been able to make a single new connection that has led to a discovery? Whereas if even a moderately intelligent person had this much stuff memorized, they would notice — Oh, this thing causes this symptom. This other thing also causes this symptom. There’s a medical cure right here.
> "Shouldn’t we be expecting that kind of stuff?"
I basically agree and think that the lack of answers to this question constitutes a real problem for people who believe that AGI is right around the corner.
Startups and AI shops: "AGI near, 5 years max" (please give us more money please)
Scientists and Academics: "AGI far, LLMs not gonna AGI"
AI Doomers: "AGI here, AI sentient, we dead"
AI Influencers: "BREAKING: AGI achieved, here's 5 things to know about o3"
Investors: stonks go down "AGI cures all diseases", stonks go up "AGI bad" (then shorts stonks)
We should stop building AGIntelligence and focus on building reasoning engines instead. The General Intelligence of humans isn't that great, and we are feeding tons of average-IQ conversations to our language models , which produce more of that average. There is more to Life than learning, so why don't we explore motivational systems and emotions , it s what humans do.
Hey, we were featured in this article! How cool is that!
> I’m not going to be like one of those spoiled children on Hackernews who could be handed a golden-egg laying goose and still spend all their time complaining about how loud its quacks are.
AGI is not going to happen. Fake it till you make it only goes so far.
The funny thing is that some people actually think they want that.
I've been confused with the AI discourse for a few years, because it seems to make assertions with strong philosophical implications for the relatively recent (Western) philosophical conversation around personal identity and consciousness.
I no longer think that this is really about what we immediately observe as our individual intellectual existence, and I don't want to criticize whatever it is these folks are talking about.
But FWIW, and in that vein, if we're really talking about artificial intelligence, i.e. "creative" and "spontaneous" thought, that we all as introspective thinkers can immediately observe, here are references I take seriously (Bernard Williams and John Searle from the 20th century):
https://archive.org/details/problemsofselfph0000will/page/n7...
https://archive.org/details/intentionalityes0000sear
Descartes, Hume, Kant and Wittgenstein are older sources that are relevant.
[edit] Clarified that Williams and Searle are 20th century.
I was thinking the same about AI in 2022 ... And I was so wrong!
https://news.ycombinator.com/item?id=33750867
It is not. There is a certain mechanism in our brain that works in the same way. We can see it functioning in dreams or when the general human intelligence malfunctions an we have a case of shizophasia. But human intelligence is more than that. We are not machines. We are souls.
This does not make current AI harmless; it is already very dangerous.
Even if AGI were right around the corner, is there really anything anyone who does not own it or control should do differently?
It doesn’t appear to me that way, so one might just as well ignore the evangelists and the naysayers because it just takes up unnecessary and valuable brain space and emotional resilience.
Deal with it if and when it gets here.
AGI by 'some definition' is a red herring. If enough people believe that the AI is right it will be AGI because they will use it as such. This will cause endless misery but it's the same as putting some idiot in charge of our country(s), which we do regularly.
Related Dwarkesh discussion from a couple of months ago:
https://news.ycombinator.com/item?id=43719280
(AGI Is Still 30 Years Away – Ege Erdil and Tamay Besiroglu | 174 points | 378 comments)
I guess using history as a guide it might be like self driving. We mostly believed it was right around the corner in 2012. Lots of impressive driving.
2025 were so close but mostly not quite human level. Another 5 years at least
What is the missing ingredient? Any commentary that dies not define these ingredients is not useful.
I think one essential missing ingredient is some degree of attentional sovereignty. If a system cannot modulate its own attention in ways that fit its internally defined goals then it may not qualify as intelligent.
Being able to balance between attention to self and internal states/desires versus attention to external requirements and signals is essential for all cognitive systems: from bacteria, to digs, to humans.
Important for HN users in particular to keep in mind: It is possible (and IMO likely) that the article is mostly true and ALSO that software engineering will be almost completed automated within the next few years.
Even the most pessimistic timelines have to account for 20-30x more compute, models trained on 10-100x more coding data, and tools very significantly more optimized for the task within 3 years
I think the timelines are more like half that. Why? The insane goldrush when people start using autonomous agents that make money.
Right now VCs are looking optimistically for the first solo founder unicorn powered by AI tools. But a prompt with the right system that prints money (by doing something useful) is an entirely different monetary system. Then everyone focuses on it and the hype 10x’s. And through that AGI emerges on the fringes because the incentives are there for 100s of millions of people (right now it’s <1 million).
We don't need AGI, whatever that is.
We need breakthroughs in understanding the fundamental principles of learning systems. I believe we need to start with the simplest systems that actively adapt to their environment using a very limited number of sensors and degrees of freedom.
Then scale up from there in sophistication, integration and hierarchy.
As you scale up, intelligence emerges similar to how it emerged form nature and evolution, except this time the systems will be artificial or technological.
The fact that no one is publishing numbers on how big their models are now is an indicative they hit a wall on training those models.
How can it be if we have yet to fully define what human intellect is or how it works? Not to mention consciousness. Machine intelligence will always be different than human intelligence.
I agree with the continual-learning deficiency, but some of that learning can be in the form of prompt updates. The saxophone example would not work for that, but the "do my taxes" example might. You tell it one year that it also needs to look at your W2 and also file for any state listed, and it adds it to the checklist.
Alan Turing had a great test (not definition) of AGI, which we seem to have forgotten. No I don't think an LLM can pass a Turing Test (at least I could break it).