Comment by zerobees
8 hours ago
It's a very brief history: it consists of three examples, only one of which isn't in the title. And the middle one is arguably a success story because the government did stop a bunch of spyware vendors it particularly disliked. That they turn a blind eye to some others is not really a policy failure, it's a deliberate political choice.
The obvious difference between PGP and Mythos is that Mythos is a service you buy from a US company, not source code you can procure from anywhere and compile on your own. So this can be enforced better without running into thorny speech issues, etc. Of course, there will be US-based actors willing to resell access, just like there are US-based actors willing to re-ship export-controlled hardware to Iran. And they will probably keep getting arrested every now and then.
The thing that's silly about this situation isn't that export controls on technology can't work, it's that it's overhyped technology and that Anthropic painted themselves into that corner by pretending it's as dangerous as nukes. Add to this a mercurial and petty administration and you have a pretty predictable outcome.
Agreed that you can export-control the closed source models pretty well (although I think the administration is gravely underestimating the long-term damage that will do to the US economy).
The bigger problem will be if someone (such as a Chinese AI lab) releases a Fable/Mythos-class open weight model. That you can't really export control successfully. Sure, you could class it under EAR or ITAR, but that's just going to make using it difficult for American companies, not everyone else. It would be a stupid protectionist measure that would only hurt the US - so I fully anticipate the admin would try it.
I can't think of anything better that could happen than a Mythos/Fable level model released with open weights. It would be a huge step forward for the whole world.
Particularly if it came paired with DeepSeek pricing.
Alternatively, we could try to keep closing barn doors after the horses are already in the next county, have been hit by a car, and are on their way to the rendering plant.
Oh, I wholeheartedly agree! My "bigger problem" statement can be read as "a bigger problem for the US government."
Isn't GLM5.2 already Mythos class?
Opus and Codex are both catching plenty of very good improvements to my GLM plans. It gets a lot right too, has a lot of good things it does, good habits and practices. But it's not as smart, not as observant, not as able to craft a nice system. In my experience.
How do you qualify what makes a model "Mythos class", and how do you reliably test for it?
1 reply →
> Mythos is a service you buy from a US company, not source code you can procure from anywhere and compile on your own
weren't chinese labs training on US Ai outputs? a looot of ai power is in correct data to train for - that's pretty much like inviting workers to your factories, they won't take machines with them, but will see and consume all the processes
I doubt that it has ever been possible to obtain enough output tokens from OpenAI or Anthropic to be useful for training other LLMs.
In any case, had that been possible in the beginning, it stopped being possible long ago, because any suspicious accounts would be banned and the cost would be prohibitive even if they were not banned.
On the other hand, anyone can train new LLMs using the open weights Chinese LLMs, or the much fewer open weights LLMs with other origins, like the NVIDIA LLMs.
So in reality it is much more plausible for a US company to use Chinese LLMs for training, than vice versa.
it is certainly possible and being done all over the place. there's a black market that chinese labs use to buy frontier american llm trajectories by the millions through US intermediaries. they're not even particularly shy about it, i have been offered $0.7 per opus 4.8 call
there's also a market for chinese labs sending checkpoints to US companies to be trained on US compute and sent back
i'm surprised that so many people take chinese tech reports about how they train their models at face value tbh
> and the cost would be prohibitive
The government of the Peoples Republic of China provides massive subsidies and incentives for R&D. The cost is absolutely not prohibitive, it's not even a factor. You are massively underestimating how much capital is involved in both countries respective industries. 500 billion on indirect compute? 好!
There’s a hugely widespread business model where criminals steal people’s API keys, then sell people (mainly in China) access to models through their proxies for lower prices, and then of course save all this data and sell it to Chinese labs for distillation.
> So in reality it is much more plausible for a US company to use Chinese LLMs for training, than vice versa.
that's... exactly the point?
make it easy to steal tech from opponent, not from you
There's no way US can keep that export control as it is for the frontier models in the long term.
This would blow huge damage to the US financial markets first, the insane CAPEX spending propping up gdp, but also US competitivity in the long run.
Sure, the US is the most important tech market on the planet, but according to Anthropic 80% of their revenue comes from outside of the US.
Let alone the fact that these research labs desperately need the top talent they can get globally, not just MIT-bred Leetcode ninjas.
The only way US can have a lead is by competing globally as it always had in the tech sector (albeit it resorted to export controls to assert its dominance over China), not by changing the rules of the game and with protectionism.
When has the USA ever shied away from destroying its own financial markets?
Honestly, a large majority of the time.
The exceptions are very notable, but it's a bad thing to bet on happening. They seem to have more problems due to protecting them too much than for the alternative extreme.
I want to put this charitably, but you come out swinging by saying this is "over hyped"... You clearly don't work in a role dealing with attacks from these models. They've changed the game, and for the worse. Capabilities that used to be available only to nation-state attackers are legitimately commodified, or nearly so.
> You clearly don't work in a role dealing with attacks from these models.
I do. And from my perspective, most people in tech do.
> these models [have] changed the game, and for the worse.
Yes. Also for the better.
> Capabilities that used to be available only to nation-state attackers
We (unhelpfully) draw a tight line around nation-state attackers. We don't perceive them as part of the relentless tendency of the powerful - to leverage tech against everyone they can. Particularly us.
> Capabilities ... are legitimately commodified,
Sure. Models will keep increasing in ability. The only choice is whether they're limited for the powerful to use against us - or if that power extends to us, to help us keep the powerful in check.
Better get used to it. Just like encryption being available to all, it was terrible for some.
Local models will catch up and then you won’t be able to legally download them. I’m guessing DeepSeek 5 will be illegal to download even if it’s only opus 4.x level when it’s released.
> Mythos is a service you buy from a US company ... So this can be enforced better without running into thorny speech issues
Do you think AWS or Microsoft would be so dominant in cloud services and office productivity if the previous administrations had banned them for non-us firms?
In my opinion no, and all across the world we would be using non-us tech.
Precisely because it's a service, the US have the interest make the rest of the world continue to send their data to them.
OK, but that's tangential to what the article is about. The article says "export controls can't work", not "export controls are not in our best interest".