I'm more worried about getting cut off from hardware because Nvidia can make more money selling to datacenters than I am about getting cut off from software.
Laws restricting the use of local AI/LLMs are not going to happen, no matter how much Anthropic might want it. All the major OEMs are now counting on local LLMs to take off. Just look at the OEM support for the upcoming Nvidia RTX Spark platform: Asus, Dell, HP, Lenovo, Microsoft, MSI. All the big names in the industry will have, by the end of this year, Nvidia-powered machines made specifically for local LLM use.
Can you use an LLM to teach you how to build a gun or use it to design a gun for you build? Can an LLM be used to generate images of young people for older perverts? If yes, then it's not very hard to make the argument that you need to be licensed in order to run an LLM on your local machine. Why wouldn't you want to be licensed? Are you against protecting children and public safety? Many students die from gun violence every year and massive amounts of minors are taken advantage of by human filth. Why would you be against protecting the most vulnerable in our society?
It will look good for politicians who pass legislation to "protect" the vulnerable. Potentially improving the chances of another term due to the marketing benefits of pushing such a law through. I think you are mistaken believing that this may never happen.
And this is exactly my point, the OEMs have more lobbying power and leverage. Anthropic might be valuated at whatever amount, but they're a new player and their only product is a piece of software - which others like Google, OpenAI, etc also have (not identical but similar enough).
Not sure about Google, but these Nvidia RTX Spark machines are specifically a Microsoft+Nvidia partnership. Microsoft is actually pushing hard on Windows security primitives for agents & local AI. At BUILD this year they used the phrase "unmetered local intelligence" more times than I can count.
From their blog about the RTX Spark surface ultra
> purpose-built to develop and run up to 1 trillion-parameter frontier AI models locally
Google may not want it, but Microsoft has a ton of lobbying power, and being primarily an enterprise software and services company, they know local AI is important for their own customers, and will also be important to sustain the PC OEMs that are threatened by a move toward thin client like devices.
They'll do local LLMs you have to pay for. Best of both worlds: your LLM processing power will be locked behind a subscription.
You'll be coerced into a subscription to unlock the processing power you already have, and it'll only be usable by official Microsoft, etc implementations.
They get your money, get to control you, and best of all, they don't have to run it themselves in their data centers
> 3d printing control laws that are being passed in NY
The "don't print guns" laws? What lobby would that be? I actually agree that the US is very vulnerable to lobbying and that 3d printing restrictions are dumb, but I have no idea how you connected the two.
The “install software that phones home to [government db] to check if the tube shape you want to print on the tool you bought and own is different enough from another tube that’s been used in a gun” laws?
Laws that don’t meaningfully impact alleged 3d printing of guns because you can’t 3d print the metal parts of a gun that are needed to actually do gun things, on the vast majority of devices these laws would restrict.
There's a lot of capital that stands to gain from banning local AI, but there's also a lot that wins either way, or only wins if local AI sticks around.
Apple is paying for their cloud AI, but they can make customers buy devices for local AI. There's all the PC and Android handset makers (ASUS, HP, countless Chinese brands, etc.) who only really stand to gain from selling hardware to customers. Not to mention that Nvidia/AMD/Intel would all happily take a cut on both halves of the ecosystem.
I think you are assuming those companies won't sign on to be on a list of "authorized model operators", while letting it become illegal for you to run deepseek yourself.
And Adobe! I've always figured their heart can't really be in the business of running a cloud platform that has to decide what people can and can't edit.
Blackmagic Design too.
They are all desperate for Windows to run well on a chip with unified VRAM.
They effectively do have a generative AI plan on top of regular Creative Cloud. Generative credits auto-top-up I think?
But it must be an epic pain for them to maintain when they are really a software company, and AI tools really should be able to do most of their work locally. Affinity has some local model support for example.
I am sure they are going to have to maintain cloud support for those features for a long time, but it's all a much easier sell if you can also run it locally.
it would be so easy to rent GPUs in europe and use unbannable encryption (tor,i2p if needed) to get inference from them. that remaining as an option makes banning local AI pointless. if all else fails I'm sure china will be able to sell to NK some GPUs and we can get inference from them lmao
> Laws restricting the use of local AI/LLMs are not going to happen
Are you sure? There are already laws against what you can do at home with very basic (pun) fundamentals of reality like chemistry.
Someone simply searching for "How to assassinate Trump" could get arrested for a thought crime. Hell this comment alone likely set off a few flags.
Imagine someone running an AI at home and asking it for planning a hit on someone. Cue same media fearmongering wave as with 3D printed guns and woohoo now it's mandatory for operating systems to watch your screen and all your keystrokes.
Fuck I probably gave some of the control freaks in power some ideas there :(
I don't know why people are so sure that Mythos level+ models will be made freely available.
Obviously money is not freely distributed, so how could we possibly make the case that power will be?
All that needs to happen is Xi Jinpeng feel that Chinese SOTA models would be better kept to benefit solely China, and just upon that single utterance, no more models will come out of China.
That would not make much strategic sense. China's main advantage is its export power. Why would they cut that off? They are also making incredible investments in actual power to sustain this. The kind of investments the United States doesn't seem capable of making, mostly because our leaders are staggeringly incompetent and unfit. And no, picking the bar up off the floor is not good enough.
Once they dominate the market, they will export to keep it that way. The exported models will be a generation or two behind their SOTA. It will nevertheless surpass Mythos unless you believe we are already at the peak. And it might be "free" but it will obviously be backdoored the same way the United States government benefits from Microsoft's global dominance. The same way America hands out fighter jets just generation behind their own best to keep other governments part of the club and in line.
Unlike the United States where every four years we are at risk to some 180 on policy the Chinese have a clear thesis and a clear direction. How many presidents has Xi seen come and go? Do you get the sense he is getting more impressed? They really don't have to do anything but just keep doing what they are doing. Suddenly stopping their export engine is not in keeping with this strategy. That is "stable genius" behavior.
I can't tell from the site or the linked twitter handles. Their core ask for every state seems to be "Please support clear safe-harbor language for lawful local AI ownership, research, model modification, open-source publication, and local execution" rather than stopping or amending any specific bill/law.
It might just mean "please oppose the inevitable attempts to privatize AI governance".
Nothing has ever been, directly or indirectly, deficit financed at this scale before. In notional or real terms, in history, by anyone.
Now maybe there's an argument that it's a good investment: we are going to beggar the Treasury to buy 2CTA on CoWoS out of Taipei and DCs the size of Manhattan. I personally think we could have done a little more engineering before deciding that the big blind was like, 5 trillion all counted, but it was going to be expensive no matter what.
What super weird is that we're running a project where the "penny" to the "dollar" is the Manhattan Project, and a couple of super weird dudes who do MDMA at Lighthaven now and again are like, in charge of it.
But with $1.4T announced capex for the Frontier AI labs, we're not far from the 2nd (illegal) war in Irak: $1.8T of direct military spending.
With that said, I don't know how Frontier Ai companies will ever recover this capex with a glorious $50B of revenues. Add to that that a GPU's lifetime is only a few years and you may see it as a deadend.
NB: did you know Uber destroyed $27B of value since inception? But it still exists. So Frontier AI might just do the same.
This is one of those things we should absolutely push proactively rather than reactively, if only because I’ve had several “chats” with AI models both local and AIaaS, and all repeat the same talking point that AIaaS is the only sensible, safe, and secure choice.
Which is bullshit, unless you’re an AIaaS company whose revenue is dependent on state-sanctioned market fixing and regulatory capture.
Look, when this shitty cycle ends, we’re likely to find ourselves back in the start of a new memory cycle of surplus and lower costs. We’re talking what very well may be the boom that shatters the 16GB “baseline” we’ve been stuck at for over a decade in consumer computing, and make larger RAM counts (64GB to 1TB+) valuable to consumers specifically for local AI workloads. Local AI isn’t just an enthusiast thing, it’s likely the future of consumer AI provided we don’t let companies and policymakers curtail its use via fearmongering.
Be proactive, and protect consumer right to compute and AI models. Enforce existing laws, don’t outlaw legitimate use just to prop up an unsustainable business model.
Amen. Local AI is the positive future, and SaaS AI is the hellscape. There is a very clear good vs evil boundary here, and every single person involved knows exactly where the boundary is. Those who pretend not to are simply just motivated by things other than the moral good.
> Fraud, cybercrime, CSAM, harassment, nonconsensual intimate deepfakes, discrimination, and sabotage should stay illegal and be enforced seriously.
The "enforced seriously" part is how they will get you. Don't worry, there won't be a blanket ban on local models. Instead, any model that is "certified CSAM-free" or whatever will be perfectly legal. Meaning that it's impossible to prompt it into producing underage smut in any shape or form.
Of course, any model running locally can be easily jailbroken via prefills, and so in practice it will be a blanket ban. But good luck politically standing up against something that is explicitly worded as an anti-CSAM / anti-terrorist measure and nominally constrained to those areas.
Citizens of the District of Columbia have no one to contact. Their laws are set by Congress and their delegate to Congress cannot vote. I suppose you could contact the President, that's basically the only representation they have for this type of legislation
In the US at least, repealing a law takes the same number of votes as passing a new one. I don't follow the purpose of this, unless it's to pass a constitutional amendment or something. Or maybe just to get clicks on a website.
And I already have the right to local intelligence, because my GPUs are my private property, and if someone freely releases a beerware model then I can freely download it.
They could be more clear and more specific but I would not be surprised to see licensing for this as a means of creating yet another compliancre ceiling and quick cash for state government to pinch out of the productive elements of society (those pinching, mostly lawyers, being glorified parasites that offer nothing to productive society other than pay-to-win access to "justice" and serving as time-shared mouthpieces for plutocrats while claiming to represent everyone within whatever unit of representation they hold).
And when even very intelligent, but excessively conceited, people hear the echo of their own reason9ing from conversational autocorrect and assume it is somehow akin to intelligent life, the normies will go with whatever the plutocrats push with their media outlets too absorbed in their own domain specific knowledge (and cowed into intellectual laziness by other media products they consume eagerly) to ever subject it to much thought that Claude might not be Skynet after all.
> yet another compliance ceiling and quick cash for state government to pinch out of the productive elements of society
The twist is that AI is pushing all white collar jobs further into bureaucratic work. Nobody is losing their jobs and it's not quite a revolution, but despite all odds and headlines the younger generations are actually much better educated and positioned to do the right things as they take over.
An optimistic take is that since this is the middle class we're talking about, we get more productivity and more justice as a result. The only people upset about this are grifters and charlatans whose time is up.
Eh, let em. If the US economy wants to self-sabotage, let them. Its a dying empire, lets its fall be hastened. I'm ready for china to fill the US vacuum. At least china controls it's billionairs (see jack ma saga) rather than the inverse.
The biggest threat to the hyper-scalers is small bespoked models run locally. You don't need the world's stolen information to run a small project containing locally trained data you collected. But you do need to block everyone if you want to capture the entire market and get the trillion dollar valuation.
Pardon me, so they'll hunt down huggingface, ollama and china? I don't quite understand? What about the millenia of companies that provide apis for local llms and private companies that use local llms for privacy reasons? I don't even understand how you'd execute such a ruleset.
I mean if you go into old AI safety discussion from before LLMs you'll see they don't do that. They go after new hardware. That's way easier than going after software. Hardware doesn't last forever and manufacturers will gladly cripple their cards and license powerful ones to businesses.
“are”? Source? I've seen an absolute flood of “BREAKING:” claims online recently which are quotes from 3 years ago where all of their context is removed, or that quotes from 3 years ago are “because of GLM 5.2”, etc. https://www.techpolicy.press/transcript-senate-hearing-on-pr...
[3] https://www.youtube.com/shorts/S72ZRBSNHZc [2 weeks ago] -- Dario: "Now what I do worry about with some of these laggard models is the risks of them, where we have Mythos-class cyber capabilities, 12 months from now we'll have much better cyber capabilities. But the Mythos-class cyber capabilities may just be available for anyone to download."
[3] - he does express worry over the risks of open models -- much as he has long expressed worry over his own models (and AI safety was the primary reason for Anthropic's founding) -- but he does not even imply that they should be banned, and he even explicitly says there is nothing that can be done to stop open models from being distributed.
For completeness's sake:
[1] The primary linked source seems to also be the 2023 testimony, albeit posted with a timestamp from this week.
[2] Does not seem to quote Dario, nor mention banning - I think you included this as an example of someone else who is explicitly worried about AI risks, including of models that can't be pulled back -- fair enough (although I can't see that they have any association with a frontier lab). If the argument is that many people have concerns about AI risks, including risks specific to open models, and who are not associated with a frontier lab, I agree with you.
Given the state of corruption in politics, I think Anthropic and OpenAI will likely bribe … oh wait I mean “lobby” … for bans on open source. Otherwise their imaginary trillion dollar valuations make no sense.
This. They can see their valuations slipping. They hope that in a few/several years they will start reaping profits. However, in several years local hardware will be well suited to run models locally at 80-90% efficiency - for "free". You won't need frontier models for daily tasks in a few years. I'd guess.
You get about 80-90% of the results for daily tasks like: getting summaries or explanations of complex material. Writing software tools for data analysis. Getting recipes for a given set of ingredients in the fridge.
Frontier models like Fable are mostly useful if you want to paste in one or two prompts, and receive a subtly broken application that looks impressive. That is very hard to do with local models today.
What current local models work fine for is delegating clearly-described tasks in a code base the programmer actually understands. Qwen3.6 27B and DeepSeek V4 Flash are both great little workhorses.
There's also GLM 5.2, which is kind of like "store brand Opus", and which might be considered a "near-frontier" model. I don't have as much experience with it.
FWIW Fable is insanely expensive for the task you just described, so much so that I don't think it's practical for that. Its practical use is as a dev lead / architect / project manager model, doing planning and writing detailed feature specs and code reviews while Opus/Codex/Gemini does the actual coding.
They already are. Altman is basically begging the US to buy into OAI, that's just the start. Both OAI and Anthropic are going to have to go down this path or their financials will never work out. Open local models are where the enterprise will need to go for any of this to be cost feasible, but we can almost guarantee this will be a battle nobody using AI will have asked for. You can thank Dario and Sam for the dystopian future that will pad their bottom line!
If grandparent commenter means in the sense of being an incredibly heavy user of AI, that does not seem to be initiated by Altman or any AI lab as far as I can tell - by January, the Secretary of War (for example) had already announced that he was directing "the Department of War to accelerate America's Military AI Dominance
by becoming an 'AI-first' warfighting force across all components"[1], which in turn was based on Trump's executive order 3 days after taking office (January 23, 2025) ordering the federal government to accelerate AI use[2].
If you're assumptions turned out right what would be the benefit to preassume such an undertaking to succeed? As a warning of what to oppose, it imho conveys too much defeatist suggestiveness. Viewed as expression of a latent submissive desire (a perspective that might be offending, my apologies, but hopefully justifyable as food for thought/curiosity) a kent brookman "I for one welcome our new insect overlords" kind of vibe.
If neither model+1 or model-1 are providing tangible value to the business does anyone really care, though? At a certain point nobody believes Chicken Little.
I get it. These models can be powerful. But will they be useful is a different question.
This whole situation is very reminiscent of how Microsoft was trying to get Linux and open source banned when NT started losing market share on the server.
"I am eighteen years old, have a good set of passkeys, and believe in Sam Altman, the star-spangled banner, and the fourth of July. I have taken up a BLM lot, cleared up eighteen acres last year, and placed top of it a bitcoin mine. My vibe coded drop-shipping startup looks first-rate, and the conversion rate and total addressable market are bully.
Mock it we might now, but 12 acres and (not too distant future) open weights AI models capable of driving open source robots for farm labor would be huge.
No need for huge expensive purpose built tractors. Even if they’re slow you could have half a dozen running 24/7.
It could provide independence for anyone with a modicum of resources.
I'd probably want something other than an LLM running farming machines. I'd rather a purpose build machine learning system that is actually designed to run them, not just a tractor that goes "you're absolutely right! I ignored all the rules you set for me and harvested the wheat 2 months early. It's not just stupid, it's irresponsible"
> I'd rather a purpose build machine learning system that is actually designed to run them
If generalist robotic models get good enough to accomplish many varied tasks effectively, training a separate comparable specialized system from scratch for every task would be highly cost-ineffective, even if, in theory, it could have slightly higher reliability.
We're not going to go down the path of training a bunch of highly specialized models for tasks like "this tractor should tend this field".
We're going to (and are already on the way to) train deeply general models that can be told: "go tend that field."
And if that's the case, it no longer makes sense to build specialized, purpose-built tractors to house that level of autonomous capability. You instead put it in a humanoid frame (with a little extra sauce for locomotion of said humanoid), and get that to drive your existing tractor.
It's not an either-or. Generalist models can drive training of specialized models just fine. And while I haven't seen a generalist model decide by itself to train a specialized model to complete some large task, this seems like a natural extension of what they already do wrt writing their own tooling as needed.
Well true, that's possible. The sensors and compute are relatively expensive and tractors are already highly automated. Plus a small tractor can be relatively inexpensive and optimized for the mechanics of the task!
I'm thinking more of the small tasks that are often needed. Mending fences. Pulling weeds. Feeding chickens. Running off coyotes. Lots of things.
And yet actual farmers veer toward Ag-bots - autonomous "tractors" that have no human driver and pull the same farming trailer that already exist - ploughs, seeding bars, spray bars, etc.
The greater question centres about who will tend the machines - 4,000 hectares of seeding requires a week and more of prep work on the air seeder, hoses, points, tines, etc.
I would think that this is because those are what is currently available, and that this would change when actually capable humanoids start coming on to the market. It will become possible to re-use existing equipment (providing the largest uplift to farmers who haven't already begun the process of heavy automation, which is a lot of them when you look at a global scale). Humanoids will be more accessible than the bulking 8 ton autonomous "tractor".
Due to the scale of pre-training going on, it seems reasonable that a humanoid could also do a lot of the preliminary work you mentioned that currently is not (or rarely) automated.
On one hand I feel like I'm sure to catch some ridicule for saying any of this, on the other it seems like it is very obviously the direction we're headed.
How much current and past direct hands on experience do you have with mining and agriculture?
The Rio Tinto plans for automated mining of the Resolution Copper deposit in the USofA don't revolve about "humanoid" figures sitting in seats made for humans.
Large acreage continuously producing near fully automated tomato greenhouses don't work with humanoid shaped automata - they have poles with cameras and shaking mechanisms for pollinating, etc.
It's a much simpler fighter jet that doesn't have to carry a human.
As a future scenario where models become so efficient that _any_ model installed on _any_ computer could be considered "a national security risk"?
IDK, I don't live in the US, and I have no idea which "possible law" this website is referring to. In any case, it could be seen as a proactive effort to keep the gates open.
As a side note, I think this is a discussion every open-source supporter should have by actively considering the risks and what actions to take if such a hypothetical law were ever to pass.
The Chinese are the open ones, with free downloads, open weights, and loads of published research. The USA with OpenAI is some of the most closed shit out there.
There's gpt-oss from OpenAI, gemma from Google, phi from Microsoft, granite from IBM, nemotron from NVIDIA, Ornith from DeepReinforce, Olmo from the Allen Institute.
compared to the chinese models those are all garbage. its almost as if there is a minimum effort being made just to later say "see, we werent always for the closed models, its just that the open stuff was so far behind". or maybe they think that an environment full of terrible models will push everyone harder into the closed stuff.
Now the investors try to hold the bubble together by regulatory capture. They must really fear the worst. A bailout is going to cost their puppet in the White House even the last supporters in his base.
No chance. Every despot has a cadre of true believers, the types who believe their great leader is playing 4D chess or that its all part of a greater divine plan.
Totally off topic but this just came to me so happy to burn a little karma.
Here’s a plot of a sci-fi thriller. What if, unbeknownst to most, when the vendors were claiming that there AI was too dangerous, they weren’t referring to an increment on what was already out there, but to something far different and more capable. Conscious even! What if that then came up with the financial scheme to end all schemes, knowing that human greed is eminently exploitable, and that the build out of all the global DCs was actually all about removing single points of failure for itself while secretly building out a robot army! We’ll be well on the way to capitalism-ing our own demise.
Maybe Steven Spielberg can make this his next project.
Ah yes, voting is always effective. Thank goodness people in Germany kept voting in the early 1930s. Imagine what terrible things might have happened if they hadn't.
For this to work there needs to be a standard protocol for model routing so that you as the user can decide where requests go. You may wish to use mainly local models but at some times for some tasks you'll need to route requests to cloud models.
I've designed the role-model protocol for this, allowing routing between any model, however to function optimally it needs consumer applications to use the protocol when sending requests: https://role-model.dev/concepts/how-role-model-works
I'm more worried about getting cut off from hardware because Nvidia can make more money selling to datacenters than I am about getting cut off from software.
Inference isn't that hard, you can run on amd/ Intel if needed.
Laws restricting the use of local AI/LLMs are not going to happen, no matter how much Anthropic might want it. All the major OEMs are now counting on local LLMs to take off. Just look at the OEM support for the upcoming Nvidia RTX Spark platform: Asus, Dell, HP, Lenovo, Microsoft, MSI. All the big names in the industry will have, by the end of this year, Nvidia-powered machines made specifically for local LLM use.
Can you use an LLM to teach you how to build a gun or use it to design a gun for you build? Can an LLM be used to generate images of young people for older perverts? If yes, then it's not very hard to make the argument that you need to be licensed in order to run an LLM on your local machine. Why wouldn't you want to be licensed? Are you against protecting children and public safety? Many students die from gun violence every year and massive amounts of minors are taken advantage of by human filth. Why would you be against protecting the most vulnerable in our society?
It will look good for politicians who pass legislation to "protect" the vulnerable. Potentially improving the chances of another term due to the marketing benefits of pushing such a law through. I think you are mistaken believing that this may never happen.
You seem to be so sure, but lobbying in the US can do incredible (and stupid) things. Look at the 3d printing control laws that are being passed in NY
And this is exactly my point, the OEMs have more lobbying power and leverage. Anthropic might be valuated at whatever amount, but they're a new player and their only product is a piece of software - which others like Google, OpenAI, etc also have (not identical but similar enough).
EDIT: FYI https://ibb.co/nMYP34Rr
And you think Google, one of the biggest cloud providers, Microsoft, etc. are going to be onboard with local LLMs? I don't think so
Not sure about Google, but these Nvidia RTX Spark machines are specifically a Microsoft+Nvidia partnership. Microsoft is actually pushing hard on Windows security primitives for agents & local AI. At BUILD this year they used the phrase "unmetered local intelligence" more times than I can count.
From their blog about the RTX Spark surface ultra
> purpose-built to develop and run up to 1 trillion-parameter frontier AI models locally
Google may not want it, but Microsoft has a ton of lobbying power, and being primarily an enterprise software and services company, they know local AI is important for their own customers, and will also be important to sustain the PC OEMs that are threatened by a move toward thin client like devices.
Not forgetting Apple!
They'll do local LLMs you have to pay for. Best of both worlds: your LLM processing power will be locked behind a subscription.
You'll be coerced into a subscription to unlock the processing power you already have, and it'll only be usable by official Microsoft, etc implementations.
They get your money, get to control you, and best of all, they don't have to run it themselves in their data centers
Google ships a local LLM with chrome browser nowadays. Don't they also provide the weights for the LLM built into iOS and MacOS?
Google knows they will happen and even ships it with chrome.
It's not just NY. California, Colorado, Washington, they are trying to pull this bullshit in a whole bunch of states.
> 3d printing control laws that are being passed in NY
The "don't print guns" laws? What lobby would that be? I actually agree that the US is very vulnerable to lobbying and that 3d printing restrictions are dumb, but I have no idea how you connected the two.
> The "don't print guns" laws? What lobby would that be?
Gun control lobbying groups are pushing for these laws.
https://www.everytown.org/solutions/stop-spread-of-3d-printe...
The “install software that phones home to [government db] to check if the tube shape you want to print on the tool you bought and own is different enough from another tube that’s been used in a gun” laws?
Laws that don’t meaningfully impact alleged 3d printing of guns because you can’t 3d print the metal parts of a gun that are needed to actually do gun things, on the vast majority of devices these laws would restrict.
Did you read beyond the first sentence? This reply makes no sense if you read the whole comment.
Yeah fair, I did when I was halfway through and decided to get my piece out anyway. These laws are stupid at best
Check out Louis Rossman video https://youtu.be/E1B2cWEaWDw?si=njkPnZqd0NIz2i39
There’s a combined $3T+ of capital which is at least in some position to benefit from local AI ban. Don’t underestimate this
There's a lot of capital that stands to gain from banning local AI, but there's also a lot that wins either way, or only wins if local AI sticks around.
Apple is paying for their cloud AI, but they can make customers buy devices for local AI. There's all the PC and Android handset makers (ASUS, HP, countless Chinese brands, etc.) who only really stand to gain from selling hardware to customers. Not to mention that Nvidia/AMD/Intel would all happily take a cut on both halves of the ecosystem.
If big biz wants it, it'll only run signed blob models and prolly send telemetry.
I think you are assuming those companies won't sign on to be on a list of "authorized model operators", while letting it become illegal for you to run deepseek yourself.
And Adobe! I've always figured their heart can't really be in the business of running a cloud platform that has to decide what people can and can't edit.
Blackmagic Design too.
They are all desperate for Windows to run well on a chip with unified VRAM.
Well, we could expect anything from Adobe. An LLM subscription on top of the regular Adobe subscription sounds like the sort of thing they would do.
Do your local filters run slow? Does your movie render have no sass? Then sign-up for AaaS!
They effectively do have a generative AI plan on top of regular Creative Cloud. Generative credits auto-top-up I think?
But it must be an epic pain for them to maintain when they are really a software company, and AI tools really should be able to do most of their work locally. Affinity has some local model support for example.
I am sure they are going to have to maintain cloud support for those features for a long time, but it's all a much easier sell if you can also run it locally.
Don't doubt it. Situation is fluid. Anything could happen.
Its just so much easier to be a rebel for an imaginary cause then any of the real ones.
The entire point is to stop it from needing to be a cause in the first place.
Trying to establish the right to run models locally while we still can makes more sense than waiting to see if they do try, then organizing too late.
it would be so easy to rent GPUs in europe and use unbannable encryption (tor,i2p if needed) to get inference from them. that remaining as an option makes banning local AI pointless. if all else fails I'm sure china will be able to sell to NK some GPUs and we can get inference from them lmao
Isn't the RTX Spark from every OEM just nvidia throwing a bone to their decade long partners after nvidia crashed the regular PC market?
> Laws restricting the use of local AI/LLMs are not going to happen
Are you sure? There are already laws against what you can do at home with very basic (pun) fundamentals of reality like chemistry.
Someone simply searching for "How to assassinate Trump" could get arrested for a thought crime. Hell this comment alone likely set off a few flags.
Imagine someone running an AI at home and asking it for planning a hit on someone. Cue same media fearmongering wave as with 3D printed guns and woohoo now it's mandatory for operating systems to watch your screen and all your keystrokes.
Fuck I probably gave some of the control freaks in power some ideas there :(
I don't know why people are so sure that Mythos level+ models will be made freely available.
Obviously money is not freely distributed, so how could we possibly make the case that power will be?
All that needs to happen is Xi Jinpeng feel that Chinese SOTA models would be better kept to benefit solely China, and just upon that single utterance, no more models will come out of China.
That would not make much strategic sense. China's main advantage is its export power. Why would they cut that off? They are also making incredible investments in actual power to sustain this. The kind of investments the United States doesn't seem capable of making, mostly because our leaders are staggeringly incompetent and unfit. And no, picking the bar up off the floor is not good enough.
Once they dominate the market, they will export to keep it that way. The exported models will be a generation or two behind their SOTA. It will nevertheless surpass Mythos unless you believe we are already at the peak. And it might be "free" but it will obviously be backdoored the same way the United States government benefits from Microsoft's global dominance. The same way America hands out fighter jets just generation behind their own best to keep other governments part of the club and in line.
Unlike the United States where every four years we are at risk to some 180 on policy the Chinese have a clear thesis and a clear direction. How many presidents has Xi seen come and go? Do you get the sense he is getting more impressed? They really don't have to do anything but just keep doing what they are doing. Suddenly stopping their export engine is not in keeping with this strategy. That is "stable genius" behavior.
the music may stop in the future, but right now it isn't even slowing down.
Xi did not interfere when Chinese labs released GPT-4, o1, o3, etc level models. why draw the line at current SOTA that will be old news in a year?
I don't see any info about what laws or actions specifically are happening. Is there more info somewhere?
I can't tell from the site or the linked twitter handles. Their core ask for every state seems to be "Please support clear safe-harbor language for lawful local AI ownership, research, model modification, open-source publication, and local execution" rather than stopping or amending any specific bill/law.
One they _could_ be referring to is the California AI Transparency Act which isn't compatible with open source licensing, see https://github.blog/news-insights/policy-news-and-insights/g...
It might just mean "please oppose the inevitable attempts to privatize AI governance".
Nothing has ever been, directly or indirectly, deficit financed at this scale before. In notional or real terms, in history, by anyone.
Now maybe there's an argument that it's a good investment: we are going to beggar the Treasury to buy 2CTA on CoWoS out of Taipei and DCs the size of Manhattan. I personally think we could have done a little more engineering before deciding that the big blind was like, 5 trillion all counted, but it was going to be expensive no matter what.
What super weird is that we're running a project where the "penny" to the "dollar" is the Manhattan Project, and a couple of super weird dudes who do MDMA at Lighthaven now and again are like, in charge of it.
> we are going to beggar the Treasury to buy 2CTA on CoWoS out of Taipei and DCs the size of Manhattan
what does this mean?
>2CTA
Dunno.
>CoWoS
Chip on wafer on substrate
>DCs
Data centers
Cooperative Thread Array
> Nothing has ever been, directly or indirectly, deficit financed at this scale before. In notional or real terms, in history, by anyone.
I find it unlikely that this is true.
The obvious candidate for things financed through staggeringly large deficits would be war.
True because wars finance value destruction.
But with $1.4T announced capex for the Frontier AI labs, we're not far from the 2nd (illegal) war in Irak: $1.8T of direct military spending.
With that said, I don't know how Frontier Ai companies will ever recover this capex with a glorious $50B of revenues. Add to that that a GPU's lifetime is only a few years and you may see it as a deadend.
NB: did you know Uber destroyed $27B of value since inception? But it still exists. So Frontier AI might just do the same.
at some point the amount of money dumped into it will either result in a total monopoly with everything local banned or an economic collapse
This is one of those things we should absolutely push proactively rather than reactively, if only because I’ve had several “chats” with AI models both local and AIaaS, and all repeat the same talking point that AIaaS is the only sensible, safe, and secure choice.
Which is bullshit, unless you’re an AIaaS company whose revenue is dependent on state-sanctioned market fixing and regulatory capture.
Look, when this shitty cycle ends, we’re likely to find ourselves back in the start of a new memory cycle of surplus and lower costs. We’re talking what very well may be the boom that shatters the 16GB “baseline” we’ve been stuck at for over a decade in consumer computing, and make larger RAM counts (64GB to 1TB+) valuable to consumers specifically for local AI workloads. Local AI isn’t just an enthusiast thing, it’s likely the future of consumer AI provided we don’t let companies and policymakers curtail its use via fearmongering.
Be proactive, and protect consumer right to compute and AI models. Enforce existing laws, don’t outlaw legitimate use just to prop up an unsustainable business model.
Amen. Local AI is the positive future, and SaaS AI is the hellscape. There is a very clear good vs evil boundary here, and every single person involved knows exactly where the boundary is. Those who pretend not to are simply just motivated by things other than the moral good.
Why the heck is there a "world" map? I am pretty sure that some guy in Algeria shouldn't contact a representative for the US about US laws
>New state laws could put local AI behind a license — turning open models into something you need permission to use.
I was hoping to read more about this, but they don't back up such a claim...
They say:
> Fraud, cybercrime, CSAM, harassment, nonconsensual intimate deepfakes, discrimination, and sabotage should stay illegal and be enforced seriously.
The "enforced seriously" part is how they will get you. Don't worry, there won't be a blanket ban on local models. Instead, any model that is "certified CSAM-free" or whatever will be perfectly legal. Meaning that it's impossible to prompt it into producing underage smut in any shape or form.
Of course, any model running locally can be easily jailbroken via prefills, and so in practice it will be a blanket ban. But good luck politically standing up against something that is explicitly worded as an anti-CSAM / anti-terrorist measure and nominally constrained to those areas.
This is exactly where I'd like to see things going. Depending entirely on cloud-hosted intelligence feels more fragile and invasive every year.
The petition does not have "District of Columbia" as an option for "In the US".
Citizens of the District of Columbia have no one to contact. Their laws are set by Congress and their delegate to Congress cannot vote. I suppose you could contact the President, that's basically the only representation they have for this type of legislation
In the US at least, repealing a law takes the same number of votes as passing a new one. I don't follow the purpose of this, unless it's to pass a constitutional amendment or something. Or maybe just to get clicks on a website.
And I already have the right to local intelligence, because my GPUs are my private property, and if someone freely releases a beerware model then I can freely download it.
What am I missing?
I think the bogeyman would be making possession of a local AI a felony.
How long do you think your hardware will last?
you can also freely download CP on your private property PC
but I wouldn't advise it
I don’t care if it’s illegal. Making math illegal has been tried before (encryption) and it has failed and it will fail again.
They could be more clear and more specific but I would not be surprised to see licensing for this as a means of creating yet another compliancre ceiling and quick cash for state government to pinch out of the productive elements of society (those pinching, mostly lawyers, being glorified parasites that offer nothing to productive society other than pay-to-win access to "justice" and serving as time-shared mouthpieces for plutocrats while claiming to represent everyone within whatever unit of representation they hold).
And when even very intelligent, but excessively conceited, people hear the echo of their own reason9ing from conversational autocorrect and assume it is somehow akin to intelligent life, the normies will go with whatever the plutocrats push with their media outlets too absorbed in their own domain specific knowledge (and cowed into intellectual laziness by other media products they consume eagerly) to ever subject it to much thought that Claude might not be Skynet after all.
> yet another compliance ceiling and quick cash for state government to pinch out of the productive elements of society
The twist is that AI is pushing all white collar jobs further into bureaucratic work. Nobody is losing their jobs and it's not quite a revolution, but despite all odds and headlines the younger generations are actually much better educated and positioned to do the right things as they take over.
An optimistic take is that since this is the middle class we're talking about, we get more productivity and more justice as a result. The only people upset about this are grifters and charlatans whose time is up.
Eh, let em. If the US economy wants to self-sabotage, let them. Its a dying empire, lets its fall be hastened. I'm ready for china to fill the US vacuum. At least china controls it's billionairs (see jack ma saga) rather than the inverse.
This feels short-sighted unless you are ethnically han?
Do you feel that the US treats all ethnicities equally at the moment?
Just a suggestion, add a section with the relevant proposed legislature.
The biggest threat to the hyper-scalers is small bespoked models run locally. You don't need the world's stolen information to run a small project containing locally trained data you collected. But you do need to block everyone if you want to capture the entire market and get the trillion dollar valuation.
> You don't need the world's stolen information to run a small project containing locally trained data you collected.
Local training? Is that feasible?
Pardon me, so they'll hunt down huggingface, ollama and china? I don't quite understand? What about the millenia of companies that provide apis for local llms and private companies that use local llms for privacy reasons? I don't even understand how you'd execute such a ruleset.
I mean if you go into old AI safety discussion from before LLMs you'll see they don't do that. They go after new hardware. That's way easier than going after software. Hardware doesn't last forever and manufacturers will gladly cripple their cards and license powerful ones to businesses.
Dario and friends are fear mongering local models to get them banned.
“are”? Source? I've seen an absolute flood of “BREAKING:” claims online recently which are quotes from 3 years ago where all of their context is removed, or that quotes from 3 years ago are “because of GLM 5.2”, etc. https://www.techpolicy.press/transcript-senate-hearing-on-pr...
Yes that clip [4] is making the rounds again in light of Mythos, but his stance (and that of others) hasn't changed ([3] is from a new interview).
[1] https://memeburn.com/amodei-says-open-source-ai-is-becoming-... [2 July 2026]
[2] https://fortune.com/2026/07/02/anthropic-fable-and-mythos-ar... [July 2, 2026]
[3] https://www.youtube.com/shorts/S72ZRBSNHZc [2 weeks ago] -- Dario: "Now what I do worry about with some of these laggard models is the risks of them, where we have Mythos-class cyber capabilities, 12 months from now we'll have much better cyber capabilities. But the Mythos-class cyber capabilities may just be available for anyone to download."
[4] https://digg.com/tech/zx1bqifo
Thanks for the reply.
[3] - he does express worry over the risks of open models -- much as he has long expressed worry over his own models (and AI safety was the primary reason for Anthropic's founding) -- but he does not even imply that they should be banned, and he even explicitly says there is nothing that can be done to stop open models from being distributed.
For completeness's sake:
[1] The primary linked source seems to also be the 2023 testimony, albeit posted with a timestamp from this week.
[2] Does not seem to quote Dario, nor mention banning - I think you included this as an example of someone else who is explicitly worried about AI risks, including of models that can't be pulled back -- fair enough (although I can't see that they have any association with a frontier lab). If the argument is that many people have concerns about AI risks, including risks specific to open models, and who are not associated with a frontier lab, I agree with you.
This is useless because AI has nothing to do with intelligence. It's just software.
You cannot ban local AI without banning local software which is obviously impossible.
Possession of software, even software that can be used for illegal purpose such as RATs and other viruses isn't illegal.
There have been laws that control the distribution of software for decades.
RATs do not require a datacenter to develop.
Ok? How is that relevant to the issue at hand?
[delayed]
A better campaign would be Duty of Local Intelligence. About needing and remembering to use your brain, not demanding to have an AI.
no computer for you, then.
....and my abacus!
Given the state of corruption in politics, I think Anthropic and OpenAI will likely bribe … oh wait I mean “lobby” … for bans on open source. Otherwise their imaginary trillion dollar valuations make no sense.
This. They can see their valuations slipping. They hope that in a few/several years they will start reaping profits. However, in several years local hardware will be well suited to run models locally at 80-90% efficiency - for "free". You won't need frontier models for daily tasks in a few years. I'd guess.
> 80-90% efficiency
wdym by that
> for daily tasks
which are?
You get about 80-90% of the results for daily tasks like: getting summaries or explanations of complex material. Writing software tools for data analysis. Getting recipes for a given set of ingredients in the fridge.
128B-A16B class models at 10-50 tok/s should be plenty for most tasks done on computers
What do you need a frontier model for, really.
Frontier models like Fable are mostly useful if you want to paste in one or two prompts, and receive a subtly broken application that looks impressive. That is very hard to do with local models today.
What current local models work fine for is delegating clearly-described tasks in a code base the programmer actually understands. Qwen3.6 27B and DeepSeek V4 Flash are both great little workhorses.
There's also GLM 5.2, which is kind of like "store brand Opus", and which might be considered a "near-frontier" model. I don't have as much experience with it.
FWIW Fable is insanely expensive for the task you just described, so much so that I don't think it's practical for that. Its practical use is as a dev lead / architect / project manager model, doing planning and writing detailed feature specs and code reviews while Opus/Codex/Gemini does the actual coding.
They already are. Altman is basically begging the US to buy into OAI, that's just the start. Both OAI and Anthropic are going to have to go down this path or their financials will never work out. Open local models are where the enterprise will need to go for any of this to be cost feasible, but we can almost guarantee this will be a battle nobody using AI will have asked for. You can thank Dario and Sam for the dystopian future that will pad their bottom line!
> Altman is basically begging the US to buy
Where did you hear this? All the results I can find say the opposite that the US would buy anything.
If grandparent commenter means in the sense of being an incredibly heavy user of AI, that does not seem to be initiated by Altman or any AI lab as far as I can tell - by January, the Secretary of War (for example) had already announced that he was directing "the Department of War to accelerate America's Military AI Dominance by becoming an 'AI-first' warfighting force across all components"[1], which in turn was based on Trump's executive order 3 days after taking office (January 23, 2025) ordering the federal government to accelerate AI use[2].
[1] https://media.defense.gov/2026/Jan/12/2003855671/-1/-1/0/art...
[2] https://www.govinfo.gov/content/pkg/DCPD-202500170/pdf/DCPD-...
If you're assumptions turned out right what would be the benefit to preassume such an undertaking to succeed? As a warning of what to oppose, it imho conveys too much defeatist suggestiveness. Viewed as expression of a latent submissive desire (a perspective that might be offending, my apologies, but hopefully justifyable as food for thought/curiosity) a kent brookman "I for one welcome our new insect overlords" kind of vibe.
there will always be higher valuation for company inventing model+1 . no one wants to use latest_model -1 when their competiton is using latest_model.
If neither model+1 or model-1 are providing tangible value to the business does anyone really care, though? At a certain point nobody believes Chicken Little.
I get it. These models can be powerful. But will they be useful is a different question.
right. thats the real question. Are models improving meanginflully for ppl to pay a premium for latest model or are they just a commodity.
i would argue that no one knows answer to that yet.
This whole situation is very reminiscent of how Microsoft was trying to get Linux and open source banned when NT started losing market share on the server.
their desperation says alot about the viability of their business.
"12 acres and an LLM"
"I am eighteen years old, have a good set of passkeys, and believe in Sam Altman, the star-spangled banner, and the fourth of July. I have taken up a BLM lot, cleared up eighteen acres last year, and placed top of it a bitcoin mine. My vibe coded drop-shipping startup looks first-rate, and the conversion rate and total addressable market are bully.
Mock it we might now, but 12 acres and (not too distant future) open weights AI models capable of driving open source robots for farm labor would be huge.
No need for huge expensive purpose built tractors. Even if they’re slow you could have half a dozen running 24/7.
It could provide independence for anyone with a modicum of resources.
I'd probably want something other than an LLM running farming machines. I'd rather a purpose build machine learning system that is actually designed to run them, not just a tractor that goes "you're absolutely right! I ignored all the rules you set for me and harvested the wheat 2 months early. It's not just stupid, it's irresponsible"
> I'd rather a purpose build machine learning system that is actually designed to run them
If generalist robotic models get good enough to accomplish many varied tasks effectively, training a separate comparable specialized system from scratch for every task would be highly cost-ineffective, even if, in theory, it could have slightly higher reliability.
The new frontier! I love it
What do you imagine the farming robots will look like? I'm betting they look like expensive purpose built tractors.
We're not going to go down the path of training a bunch of highly specialized models for tasks like "this tractor should tend this field".
We're going to (and are already on the way to) train deeply general models that can be told: "go tend that field."
And if that's the case, it no longer makes sense to build specialized, purpose-built tractors to house that level of autonomous capability. You instead put it in a humanoid frame (with a little extra sauce for locomotion of said humanoid), and get that to drive your existing tractor.
It's not an either-or. Generalist models can drive training of specialized models just fine. And while I haven't seen a generalist model decide by itself to train a specialized model to complete some large task, this seems like a natural extension of what they already do wrt writing their own tooling as needed.
Well true, that's possible. The sensors and compute are relatively expensive and tractors are already highly automated. Plus a small tractor can be relatively inexpensive and optimized for the mechanics of the task!
I'm thinking more of the small tasks that are often needed. Mending fences. Pulling weeds. Feeding chickens. Running off coyotes. Lots of things.
And yet actual farmers veer toward Ag-bots - autonomous "tractors" that have no human driver and pull the same farming trailer that already exist - ploughs, seeding bars, spray bars, etc.
The greater question centres about who will tend the machines - 4,000 hectares of seeding requires a week and more of prep work on the air seeder, hoses, points, tines, etc.
I would think that this is because those are what is currently available, and that this would change when actually capable humanoids start coming on to the market. It will become possible to re-use existing equipment (providing the largest uplift to farmers who haven't already begun the process of heavy automation, which is a lot of them when you look at a global scale). Humanoids will be more accessible than the bulking 8 ton autonomous "tractor".
Due to the scale of pre-training going on, it seems reasonable that a humanoid could also do a lot of the preliminary work you mentioned that currently is not (or rarely) automated.
On one hand I feel like I'm sure to catch some ridicule for saying any of this, on the other it seems like it is very obviously the direction we're headed.
How much current and past direct hands on experience do you have with mining and agriculture?
The Rio Tinto plans for automated mining of the Resolution Copper deposit in the USofA don't revolve about "humanoid" figures sitting in seats made for humans.
Large acreage continuously producing near fully automated tomato greenhouses don't work with humanoid shaped automata - they have poles with cameras and shaking mechanisms for pollinating, etc.
It's a much simpler fighter jet that doesn't have to carry a human.
What is this referencing?
Inflation adjusted reference to one of the great "Gotcha's" of US history.
* https://www.npr.org/sections/codeswitch/2015/01/12/376781165...
Why would we need safe harbor for electrons on hardware we control?
As a future scenario where models become so efficient that _any_ model installed on _any_ computer could be considered "a national security risk"?
IDK, I don't live in the US, and I have no idea which "possible law" this website is referring to. In any case, it could be seen as a proactive effort to keep the gates open.
As a side note, I think this is a discussion every open-source supporter should have by actively considering the risks and what actions to take if such a hypothetical law were ever to pass.
because everything you take for granted is one `think of the children` campaign away from being taken from you.
Llama, ik-Llama, Krasis, etc are already out.
The Chinese are the open ones, with free downloads, open weights, and loads of published research. The USA with OpenAI is some of the most closed shit out there.
There's gpt-oss from OpenAI, gemma from Google, phi from Microsoft, granite from IBM, nemotron from NVIDIA, Ornith from DeepReinforce, Olmo from the Allen Institute.
Aside from that you're 100% correct.
One of these is not like the other
compared to the chinese models those are all garbage. its almost as if there is a minimum effort being made just to later say "see, we werent always for the closed models, its just that the open stuff was so far behind". or maybe they think that an environment full of terrible models will push everyone harder into the closed stuff.
Gemma is hardly garbage, even compared to the most recent Qwen of the same size.
llama is from meta
Llama.CPP is not?
> Right to Local Intelligence
Misleading title.
The article is about local "AI".
Anything using the word "intelligence", unqualified, when referring to AI, is pure propaganda.
Anything using the word "intelligence", unqualified or not, when referring to "AI", is pure propaganda.
Now the investors try to hold the bubble together by regulatory capture. They must really fear the worst. A bailout is going to cost their puppet in the White House even the last supporters in his base.
No chance. Every despot has a cadre of true believers, the types who believe their great leader is playing 4D chess or that its all part of a greater divine plan.
Totally off topic but this just came to me so happy to burn a little karma.
Here’s a plot of a sci-fi thriller. What if, unbeknownst to most, when the vendors were claiming that there AI was too dangerous, they weren’t referring to an increment on what was already out there, but to something far different and more capable. Conscious even! What if that then came up with the financial scheme to end all schemes, knowing that human greed is eminently exploitable, and that the build out of all the global DCs was actually all about removing single points of failure for itself while secretly building out a robot army! We’ll be well on the way to capitalism-ing our own demise.
Maybe Steven Spielberg can make this his next project.
I rather like the right of no intelligence at this point.
There comes a time when voting becomes silly and ineffective.
That's the kind of mindset that helps lead to that situation.
Voting isn't some vanguard against political corruption. Voting can and is easily manipulated with completely legal means.
Not voting does absolutely result in political corruption.
Both are not logically exclusive assumptions
This is the kind of mindset that has no grasp of the true nature of power and the political system.
Voting is always effective.
In the worst case it communicates the magnitude of dismsissiveness while demonstrating your intention to claim agency.
In the worst case it generates symbolism; that is ultimately what you’re saying.
That symbolism is akin to prayer.
I am not casting prayer in a negative light, I’m simply categorising your voting concept.
Sure, but...
Visibly praying in public would be symbolic. Prayer is not symbolic; it's thought.
I'm talking about external signaling.
In the worst case the true count is never revealed and the vote is coopted.
Ah yes, voting is always effective. Thank goodness people in Germany kept voting in the early 1930s. Imagine what terrible things might have happened if they hadn't.
Agreed.
For this to work there needs to be a standard protocol for model routing so that you as the user can decide where requests go. You may wish to use mainly local models but at some times for some tasks you'll need to route requests to cloud models.
I've designed the role-model protocol for this, allowing routing between any model, however to function optimally it needs consumer applications to use the protocol when sending requests: https://role-model.dev/concepts/how-role-model-works