Comment by jmward01
10 days ago
This is regulatory capture in action. This will make it hard/impossible for new vendors to come into the market and only established companies will get to play, and charge, for LLMs. What does this mean for open source? Will it become illegal to download weights? What about train your own? Are we heading to a world where GPU use is regulated to ensure that illegal LLMs aren't being processed on your machine? More broadly though, how will this stop anyone but average people? Countries outside the us will completely ignore this and keep developing and moving ahead. Maybe Europe will adopt similar things but the genie is out. I can train insainly powerful models on my laptop. If you want to stop LLMs with legislation you can't do it like this.
As far as Europe is concerned they have recently signed up to the 'Pax Silica'[1] and willingly givrn the LLM space over to the US incumbents with buildtin legislation banning Chinese models and coperation with them. So EU will be a renter of the LLMs that the US allows them to use. In the long run OpenSource will dominate as it did in the DB(MySQL/Postgres)/ServerOS(Linux/BSDs) versus Proprietery rent seeking alts like Oracle and Microsoft et al. Would be interesting to see what the global startups using Qwen/DS/Kimi etc within the EU-US space navigate the cutting edge OpenSource LLMs vs seeking/getting a permission slip from the US gov.
[1]https://archive.is/aiJiq
I hope that open models will dominate. The difficult part to reconcile for me is the amount of compute that's required to create and run such models. Small models are fine, I run local llms 27b param on a gpu, but it's not even close to frontier in capability. Who wants to drop $40k+ on hardware to run these things. Companies, maybe/perhapts. On the other hand, to run a DB I can get a server for $3k and handle tons of traffic on it and other things too.
Has anyone tried to run a data center as a Co-Op?
11 replies →
I believe until the hardware designs catch up to be more commodized ala cryto mining evolution from GPUs to ASICS for specfic algos. Designs (like Google TPUs equivalent) would also need to evolve to be more memory dense to be able to handle them. Untill then it seems will be system time shares for the larger models , probably with a bring your own model and pay as you go.
5 replies →
There might be a community effort at some point. This happened in chess where the community recreated and then improved on Alpha Zero. You could run small training chunks on your machine. Some people donated thousands of hours of server time.
They never will. The only reason China releases open weights is because they can't compete on frontier models. Whoever has the frontier model has no incentive to give it away for free.
They have signed a non-binding agreement to potentially cooperate on AI supply chains. It's hardly a declaration of fealty, nor does it have any practical impact on the use of Chinese models in the near term. I'd view this more as hedging their bets for the future.
> It's hardly a declaration of fealty
As a European, the way I see it, Europe declared fealty to the US and relinquished its sovereignty a long time ago, sadly.
Also, the way these agreements tend to work is that you agree that you won't source from the 'enemy side' i.e. China. It works this way for NATO and it will work the same way here.
17 replies →
This is beyond ridiculous.
At the same historical turning point when Europeans are finally waking up to the need to become less dependent on their so called US ally for weapons production and security, they are immediately choosing to become dependent on the next layer of critical infrastructure.
Instead of learning the obvious lesson, Europe seems ready to purchase the future from whoever Washington allows it to purchase from. It used to be the guns, now it's the AI.
It is so idiotic and short sighted that you can barely even blame anyone who keeps exploiting this over and over again. It is always the same story.
Europe has been burning furniture to keep the house warm for so long that an entire generation now thinks that chopping wood is for suckers.
2 replies →
EU is very lucky USA still has very strict immigration policy. If USA made it easy for highly skilled professionals to emigrate there then there would be almost no one capable left here.
2 replies →
big misconception
open source != open weights
open weights model is like... Winamp for example. it's free, you can download it and use it however you like, you could also do some binary patching or dll injections to alter it functionality but it's not enough to develop next version.
the same is with ai models, weights are the binary final artifacts. for development and improvements you need to have training data, pipelines, RL harnesses, etc.
also of you believe Chinese companies will be releasing weights indefinitely, you are not understanding motivations.
Chinese companies spend significant amounts of money to train a model so why they are releasing it for free? they basically provide researchers starting point for developing tooling and optimizations for serving the model in return. and also get some PR. They also do not have to pay for inference of those models that much, as they probably serve them with loss anyways to gain market. they are gov sponsored so money are not issue there, so they try to speedrun their way to what US companies have. And guess what happens when they reach it. they will stop releasing weights and increase pricing or will use them for gov purposes.
It doesn’t look like "the EU" has signed anything, just that a couple EU countries independently accepted to join the summit.
europe2031.ai
Was looking for this. It’s a great short story (it starts with real events and then predicts the future) leading to European fealty to US . Seems increasingly plausible.
13 replies →
> [Washington] piles pressure on The Hague to halt ASML’s remaining exports and servicing of its DUVi machines.
> The Commission backs the Dutch [...]. The European position fragments before it has properly formed.
That the EU would, after recognizing AI's value, freely give up control of its few advantages to the US, despite of this being a conventional trade issue which the EU has experience with, seems like a very pessimistic assumption.
(I stopped reading after that part.)
Germany and Netherlands signed alone, not as EU block.
[dead]
> I can train insainly powerful models on my laptop.
What's with this hallucination? The thread is about GPT-5.6. Your laptop can't even run gemma 4 unquantized bfloat16, which is light years behind GPT-5.6, and running it is light years from training it. If something that a laptop can train is insanely powerful for you, you don't need to worry about this thread at all.
Indeed every AI enthusiast has to find within their heart their consistent position between the two extremes:
- Regulation like this is dangerous because the accessibility and capability of open-source AI training and inference are limited;
- Regulation like this is hollow because the accessibility and capability of open-source AI training and inference are unlimited.
Regulation like this is dangerous for the US because it can kneecap our country, but I don’t think it’ll affect the pace of innovation globally. China isn’t going to slow down just because the Trump administration asked for some “ballroom donations” or whatever nonsense is going on behind the scenes here.
Eh, they said models, which can mean hyperspecific domains, not necessarily a massive generalized LLM.
Has the US government limited access to any non-LLM ML models?
This also completely screws over U.S. businesses. American startups will be forced to pay premium prices for nerfed, heavily-censored, 'compliant' models from a few massive corporations. Meanwhile, foreign startups will be running cheaper, unrestricted open-source AI. We'll price ourselves right out of the global market.
What government wants to have their population use foreign AI. (Not many). Only issue I see is good enough is what the majority might be okay with.
This is probably not the intent.
US intelligence has gamed this out and sees risks at the fringe - advanced artificial intelligence moves the needle the most in countries economically & politically dominated by the USA, the Global South are the targets of this, not US startups.
So far it's only US doing this. I don't think it's in anyone else's interest to limit development of open source models or chips. Nvidia has secured a leading position in GPU market by being the best overall, but if US continues to mess up with the export, that changes the calculation and surely we'll see the alternatives
Any country that developed sufficiently advanced models will pursue the same path. The EU as a regulatory body would pursue the same path, as will/is China.
It’s rather straightforward to think through. If China (they’re practically the only competitor) built a sufficiently advanced AI system would they allow it to propagate on the free market? Of course not. The fact of the matter is that they are behind, even if it’s just a little bit, so the best strategy they have is to try and compete with “good enough” models with lower/subsidized cost - but that is a losing strategy because AI is and always has been a winner-take-all strategy.
Likewise if, idk, France someone built an AI system that was valuable do you think they’d just hand it over for sworn enemy Donald J. Trump to utilized? Of course not.
The American strategy in the context of the current geopolitical landscape is the only valid one and the obvious one. If you find yourself criticizing the American strategy it’s because you aren’t in the arena where you would, inevitably, make the same decision to restrict access.
This isn't how things typically work. For instance the US is increasingly adversarial towards China yet China continues to largely power the US economy both through manufacturing and market access, which makes up an increasingly large share of all revenue for many US companies. Why? Because that position as a dependency is not only directly economically beneficial to themselves, but also provides leverage which can be utilized in extreme circumstances.
This isn't really going to achieve much besides incentivizing the growth of non-US models and minimizing market access for US-based models. I was expecting the US government to try to ban foreign models, which is also a self-own but orders of magnitude less than this. All this will do is greatly diminish the influence of the US in the future, and minimize the benefits they might reap from a global LLM explosion. It'd be like if 30 years ago, China decided that their manufacturing could only be used by white-listed individuals. Their economy and influence wouldn't be even a fraction of what it is today.
1 reply →
> AI is and always has been a winner-take-all strategy
People claim that AGI is. AI is turning out to be a fairly competitive but “normal” product. Companies carving out niches on cost, quality, and speed.
If it was a winner take all OpenAI’s head start would have been decisive. For years ChatGPT was far ahead of everyone, then Anthropic released Opus 3, then OpenAI released 4o, then in mid 2025 it seemed like everyone had strong reasoning models including Google with Gemini 2.5, and now Claude is probably the best coding model. So taking the top spot is not a guarantee you can hold it.
Also the top model becomes a prime target for distillation, making it easier for competitors to keep up.
4 replies →
I think you view this situation from the US point of view and assume that China has the same guiding principles and values in their foreign policy, for which it doesn't. They might do what you said, of course. But they very well might also treat LLMs as another goodwill investment like the Belt and Road Initiative (https://en.wikipedia.org/wiki/Belt_and_Road_Initiative) and export the capability to partner countries, for example, in Africa, to strengthen relationships.
16 replies →
> AI is and always has been a winner-take-all strategy
AGI sure. But I don't think we're going to get there in our lifetimes, if ever. There are too many structural and physical limitations. One everyone seems to be catching onto now is that they're running out of human data and are incestuously feeing AI output to itself as input.
Current state AI is barely an improvement over where it was fifty years ago. We just have stronger hardware and more content to train on. We need a new paradigm. One that hasn't come in half a century.
> Any country that developed sufficiently advanced models will pursue the same path.
Looking at most of the available evidence, Mythos is an incremental upgrade over other models and nowhere near the implied advancement that this seems to point to. I guess you could be right in that a sufficient advancement would cause this type of withholding of it, but I think it's kind of silly to think that the US has reached that level.
I don’t think the ai market has been around long enough to prove winner take all, I see it more like a commodity which is why the winner quickly switched from OpenAI to Anthropic at least on coding.
I generally find the Chinese models to be superior. If you look at the token pricing or hosting cost it’s still about 80% margin on Chinese models. Deep cuts should still be possible.
the only people its relevant for is the people in first. We wont know what any other state would do until someone passes the us, if that happens.
it sucks that we're in a place where the us has an dishonest leadership, because the current situation would be pretty reasonable if any other admin was in charge.
let models go free, until one proves dangerous in the real world then require gov approval after that.
I don't think anyone rational would have the position everyone should have insta access at the same time to the highest model once it crosses the point of enabling actual dangerous things.
> This will make it hard/impossible for new vendors to come into the market and only established companies will get to play, and charge, for LLMs.
I don't understand how you leap from "US govt. decides who gets to use GPT-5" to "limit new vendors from competing with OpenAI".
Can you walk me through the logic?
> the U.S. government would initially approve who gets access to its latest new release while AI companies and the administration work out a longer-term plan for regulation on the sector.
So you're a new vendor with a GPT-5.6 or Mythos class model. How do you suppose the regulations are going to work? First you need to get on the list of companies that are allowed to release models, and then you need to have a whole system for limiting access. Both are going to be hugely expensive, on top of training new models already being insanely expensive to begin with.
Thus, it's not a legal limit, but a real practical limit because it's too expensive. If only OpenAI and Anthropic and Google can afford to jump through the hoops, they've effectively gotten to "limit new vendors from competing with OpenAI".
> So you're a new vendor with a GPT-5.6 or Mythos class model. How do you suppose the regulations are going to work? First you need to get on the list of companies that are allowed to release models, and then you need to have a whole system for limiting access. Both are going to be hugely expensive, on top of training new models already being insanely expensive to begin with.
So your thesis is that someone has the resources to create a model at the level of Mythos or GPT 5.6 but bot have the resources to jump through the concomitant legal requirements?
Surely that’s unlikely?
1 reply →
These frontier models are now good enough that they can assist heavily with new optimizations for future models, including, code them. Restricting their usage to a few companies takes away that advantage away from other companies, thereby, limiting new vendors from competing with OpenAI.
> Restricting their usage to a few companies takes away that advantage away from other companies, thereby, limiting new vendors from competing with OpenAI.
wait, are you saying a competitor needs access to an OpenAI model in order to build a competing model?
5 replies →
Do you have any sources for the claim that LLMs meaningfully help in the production of LLMs?
6 replies →
>This is regulatory capture in action.
Isnt this all export control based? If so its not regulatory capture for a few reasons. If not disregard this.
1) new entrants wont get export controlled because they arent leading edge
2) a new company could just implement KYC. It could even be a competitive advantage (Anthropic wont or cant)
> Only companies approved by the government will get access. There is no process for individual users to get access to the new model.
That sounds like more than just export control.
This means that big company A that the president has a business interest in could get access to the most powerful AIs, while a startup competing with it doesn't.
It's worth keeping in mind the broader context here politically. The president and his cabinet are generally pursuing policies of a strong unitary executive branch that centralizes "technocratic" functions under the control of politicians and political appointees. The majority in Congress doesn't appear to mind and seems to be actively sitting back from legislating on topics like this. The overall effect is that of creating a system in which the rules are deliberately uncertain and the only reliable way to get approval is by aligning yourself or your organization politically. It's a powerful technique for ensuring political compliance in the corporate world.
2 replies →
Of course. The US government inserts itself into commerce, starts picking winners and losers, and thus creates an industry for petitioning gatekeepers at all levels for access. The goodfellas profit and the public pays. OpenAI and company get manufactured scarcity to raise margins, and lawmakers get to collect an entry fee. This was doomed to happen. The US is much closer to a total state than most give it credit for.
I guess if the new model has the capability to do 'something' that's national security threat, then this makes sense. Otherwise this move makes zero sense, and actually is a drag on innovation - who wants to invest money and people to make a better model that can't actually be sufficiently sold to make a return on investment? Meanwhile other models from Europe or China that are better steal market share. Though that's not to say they won't do the same thing for the same reasons.
> Are we heading to a world where GPU use is regulated
Well, there was the kerfuffle around PS3 (IIRC) and 'supercomputers'. I suppose that would be the 2020s version of that. What's old is new again. Ultimately you can just continually wire together less powerful hardware to come up with something that can do the job.
>Well, there was the kerfuffle around PS3 (IIRC) and 'supercomputers'.
There were no export restrictions on the PS3.
They were probably thinking of the PS2, which Japan did, in fact, temporarily restrict: https://www.latimes.com/archives/la-xpm-2000-apr-17-fi-20482...
They meant PS2. Japan was worried the processor was powerful enough to control missiles and put restrictions on how many can be sold to a single person and banned certain countries.
1 reply →
Double precision cell chips were reserved for military, I think (or maybe also blade processors). So doing any serious physics on them was dead in the water - same with macs tbh
1 reply →
I said kerfuffle. I didn’t say ‘export’.
Is referring to these new models as "LLM"s still correct anymore? The frontier models are a more complex orchestration of LLMs and many other programming techniques, they aren't just a weight set anymore.
The whole purpose of this is to let non-technical users say things like "make a flappy bird games" and have it succeed after using 10 million tokens. A decent engineer could build the same thing in 20 mins with a small local LLM.
Same thing with a cyberattack, someone who generally knew what they were doing could quickly surface the info to perform each step of an attack. The govt just wants to ensure random 12 year olds can't type: "how I hack fortnite?" and get a step by step guide.
I think it correct to refer to the models as LLMs. The references I can find for the techniques used seem to boil down to transformers + optimizations to make transformer-index larger/faster + chains-of-thought-optimization/the ability to add real time and other data to the context window.
See: https://kenhuangus.substack.com/p/frontier-ai-models-technic...
This isn’t about keeping people from having the power of frontier LLMs. So tricks that let others have it aren’t a defeat of this policy.
This is export control, where the US government seeks to leverage the fact that these frontier models are US made. This is then leveraged against opponents, and likely also just for grift. There’s also perhaps a little legitimate worry about the implications of free access to, but that is secondary to the real goal.
I think in the long run this will severely hurt commercial models since they cannot reliably used in workflows and you don't want to be dependent on some trigger-happy random guy.
It is free advertising for now, but in my circle people are looking for alternatives and are surprised how viable they have become.
> I can train insainly powerful models on my laptop.
This is such hyperbole. You might be able to train a model that's merely useful in a single domain, but to say you can train an 'insanely powerful' model on consumer hardware is laughable.
Zuckerberg has spent billions trying to keep Meta's models relevant, somewhat unsuccessfully.
If only he had known he could simply use his laptop to train them.
Warm up your VPN to zAI for the eventual banned GLM-6 I guess
what make you believe that the Chinese gov is going to allow GLM-6 to be made publicly available?
Let me frame it like this;
If GLM-6 is made publicly available[1], will that change your predictions of how they will behave in the future, or your understanding of their motivations in the present?
[1] I am certainly no expert of Chinese AI policy, but I fear you are attributing US values and goals to the Chinese govt.
From where I stand, they appear to be very different though. China is seeking to increase influence, whereas the US (seems to be) seeking to become more insular. China is actively persuing the world-leader role, while the US dismantles soft-power tools like USaid, and alienates countries pursuing obviously flawed military excursions.
Why would the Chinese gov...spoil their rival's tulip bubble by dumping even fancier free tulips on the market?
What? This is the opposite of regulatory capture. Neither anthropic nor openai are getting to choose what happens with their models.
Regulatory capture doesn't necessarily mean the regulated get to decide what the regulators do in precise steps. It can simply mean they support and exist within a regulatory regime that greatly benefits the regulated.
In fact, you generally don't want them directly telling the regulators what to do. Instead, the regulators make complex, costly rules that only large establishment players can follow. The regulators look like they're doing their job; the regulated enjoy higher margins and protection from disruption.
How does this benefit the regulated? I'd say it dooms the regulated:
* with the government requiring the regulated to obtain approval to add each customer, they're losing a massive number of their customers
* and even if a customer gets approved, every such customer now sees that access to the regulated can (and has been) shutoff with no notice if the gov doesn’t like the provider or customer - it's now a massive supply chain risk for any customer to use a regulated provider
* the regulated losing a massive part of their customer base for both of the above reasons means significant impairment of their revenue, as well as their valuation, and staying ahead of their competitors and open models requires massive ongoing investment
Open models are mere months behind.
1 reply →
And the regulated may even publicly complain about the regulations, to increase the illusion that it isn't regulatory capture.
That's just vanilla regulation. Yes it tends to create market inefficiencies, by definition. Capture entails industry corrupting and directing regulatory bodies.
Exactly! The thing that squeezes out new entrants isn't only the compliance cost, it's that your whole roadmap ends up resting on access you don't control.
We already saw the terms moved under us once with Fable, the retention policy changed and some requests started routing to a weaker model, none of us small operators had any say in that. Now access itself is a government decisions.
For anyone building on top of these APIs that's the real barrier, not the rule-following overhead but the fact that the ground can shift mid-flight and you can't negotiate with whoever's moving it.
Which is exactly why open weights start looking less like ideology and more like risk management.
The other answers are also valid, but lest we not forget, Sam is openly a fascist supporter and is clearly in bed with the regime in that he funded 47's campaign and jumped in to rescue Hegseth's automated kill list with OpenAI's GPT when Anthropic refused. Furthermore they are likely operating on some kind of quid pro quo agreement even if it's not public knowledge, because that's how all this bribery stuff works. Bezos agreed to use his media empire including WaPo to spout MAGA propaganda, for example. It's trump's one and only MO so to assume it doesn't apply here would be insane.
So, while OpenAI may not in a legal/technical sense, be the benefactory, that is not required for the term to apply, AND they may as well be considered party to the creation of the regulation since they have openly lobbied for it, openly inserted themselves into the government apparatus both formally and informally, and likely are co-conspirators to whatever Trump's autocratic self-enrichment scheme is.
I think you meant to post this on Reddit
1 reply →
As I understood they will limit running those undesired llms with Remote Attestation and Trusted Execution Environments.
How is this regulatory capture? Any new LLM company can just exist outside of the US and export to everyone.
You’re saying it’s not regulatory capture because it’s not enforced globally?
it's not regulatory capture, because they are not regulating what customers can use - it's limiting what some companies can serve. it's way less impactful to the market as a whole.
Until they can’t.
It is as if people can't imagine the concept of sanctions despite their liberal use over the years.
1 reply →
Right. That's what sanctions can accomplish...
Countries who've made the mistake of allying with the US might face sanctions or some sort of threats. People will just use Chinese AI then. This is the US biting itself in the ass.
Nick Bostrom wondered aloud in Superintelligence (2014) why states would allow individuals and private organizations to develop AGI. If one takes the possibility seriously, AGI would a source of immense power and any state would to take that opportunity for itself.
Edit: Not saying whether AGI is right around the corner, that's a different discussion. I'm just saying that a serious possibility of AGI and an understanding of possible consequences will make a state act.
Well states are made of individuals, and at least in the US we should be able to self determine via elections and public discourse..
If that's impossible in any meaningful way, then yes, doesn't matter which color jersey the government is wearing, it's authoritarian.
AGI is religion invented for the stupid. There's no world in which a silly LLM can make intelligent decisions. It's all smoke and mirrors to make insane amounts of money and to maintain the sheep agreeing with the powerful.
What rock have you been living under?
>can train insainly powerful models on my laptop
Can you give an example or two?
> I can train insainly powerful models on my laptop
Can you?
Correction: Can you yet?
Anthropic and OpenAI: Dear Uncle Sam, it seems laptops became too powerful, can you please do something
Uncle Sam: For national security reasons starting from now on every purchase of GPU model with higher than X Petaflops will need written permission by the US president
Anthropic and OpenAI: Look poor citizens, we are willing to share our capacity with you in limited form, by using our LLM you can avoid spending 35 years in the waiting list to buy a GPU, by the way, to simplify pricing, here is our new pricing with 5700% increase. Enjoy
6 replies →
Extrapolating the progress in both hardware and model efficiency, that will take decades
Depends on the point of view, I suppose.
Powerful enough to shock someone in 2010 with a wikipedia chat bot? Possibly.
Powerful enough to shock jaded HN commentators right now? Possibly not.
its not about whether you can shock anyone, if anyone is driving cars outside, you can't say you have SOTA-horse to go from A to B.
When models are good, expectations are adjusted accordingly to deliver things on par with the whole industry, you can't just say, I have built my own Intel Pentium II, now I will try to use it to compile Electron App and run 3DS Max there
5 replies →
literally the opposite of regulatory capture - it spells trouble for specifically openai and anthropic mostly
That's often how regulatory capture works. It seems bad for the incumbents, but it's worse for any future competitor. As long as the cost of complying with the regulation is less than your expected loss from competition entering the market, then it's net positive for the incumbent.
I think this is plain old regulation, not regulatory capture. For it to be the latter we'd need evidence than it was the incumbents who designed this oversight which they then asked the white house to rubber stamp.
1 reply →
GPUs and computer hardware prices have been on the rise. I can see a twisted perspective where it’s stated that the US government needs to closely control computer hardware that can run particular LLMs, as a national security interest. At least that idea isn’t completely wild now seeing what we have been experiencing.
Remember those weird conspiracies we used to have about universal surveillance; tracking and so forth? Well if you think back to those and whatever might happen with GPUs and hardware, and LLM restrictions or the likely age gating//ID’ing that is to come from this, that’s a good guiding framework for how this will proceed and affect normal people.
>I can train insainly powerful models on my laptop
Explain. The labs have been spending about $100 million in compute to train a model.
Give me three ish months and, I hope, to show exactly what I mean!
Please go ahead. If you are right, you can raise billions of dollars for your startup.
I think I understand now. Don't do it for me. In fact, I'd rather you found something else to do.
> This is regulatory capture in action.
With the twist that it will end up involving payments that directly benefit Trump, following the Mafia business model that he learned in the construction industry and that he's brought to the White House.
Buy Bitcoin.
look, i'm sorry, but this is a thoroughly solved problem by now.
How does owning some bitcoins let you run a powerful LLM?
I take it you've never known anyone who has tried to run even a small consumer AI company. Do you think that they all independently come up with the same censorship rules, despite no law mandating them? Why do you think that is? That they are just that well-thought out that everyone happens to agree with that same intersecting boundary?
8 replies →
> Are we heading to a world where GPU use is regulated to ensure that illegal LLMs aren't being processed on your machine?
Actually that sounds pretty reasonable considering current regulations regarding almost any other important resource / material that affect the general population.
You Gen-Z people serously need what Free software, Free society meant. We fought against people like you and against treacherous computing. And we will do again.
I'm 43 years old, dude. Stop fighting with people and learn to adapt.