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.
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.
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
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.
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.
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.
It is possible to shutdown access to hosted services, as happened with Fable, but it can't really be done selectively. The US government wanted to allow it for US nationals only but Anthropic couldn't do that and so shut it down for everyone. Even if they did tie Claude accounts to nationality some people would set up "proxy servers" to allow access (either for montetary gain or because they don't agree with the restrictions)
No, that actually means something went well - my adblocker saved me from being blasted by distracting, deceitful, dangerous content.
Pro tip: reader mode bypasses this aggressive “go away” banner.
Huawei is a good case in point, about to have a 100% domestic replacement for NVIDIA chips (& CUDA stack), not reliant on TSMC, ASML, Samsung, SK Hynix... Initially Huawei's Ascend AI chips had used HBM memory from Samsung and SK Hynix, but their next generation 950 series (fabbed by SMIC) will use memory (fabbed by CXMT), not using the HBM standard, but by necessity their own HiBL and HiZQ standards.
HBM depends on ever wider memory bus widths to increase bandwidth, which in turn depends on SOTA TSMC manufacturing nodes for bus density. Huawei found a different way, using their LinqQu interconnect/switching tech to aggregate the bandwidth of individually slower memory chips resulting in an aggregate 4TB/sec bandwidth on par with HBM3e.
Trump has blocked Fable for export, but China (Ziphu) already has GLM-5.2 knocking on the door of the US frontier models, despite being developed with one hand tied behind their back. GLM-5.1 had only scored 18% on DeepSWE, but GLM-5.2 coming 11 weeks later, scores 48%, about on par with GPT 5.5. What's next ?!
DeepSeek and Qwen would have been neat parlor tricks, never developed further because the US subsidized mega models would have been too cheap and reliable to make them necessary.
Nope! USA played its hand and forced the world to have to account for its USA exposure risk and deversify. A serious one time use strategic advantage pissed away.
Doesnt really bother me personally per se; but armchair quarterbacking the USA's decisions seem Suboptimal...
History shows that export controls fail on knowledge, but are damn effective on commercial products.
There's no way to justify high capex, and US-personnel only research labs and pretend to keep the lead in AI at the same time.
You either compete square and fair, or you're gonna stay behind.
Doing god’s work there, Andy, thanks /s
Wonder what Anthropic internal messages look like about his move. Does Anthropic have a meme slack channel?
I don’t know how much we want to believe the “reports”. But there are probably a few other tricks they didn’t expose. If these are pre/post processing guardrails I could see something like “fix bugs” actually working.
This was never the question. The question was, will the export controls slow them down in the short/medium term to the point where it will give US companies an advantage? And the answer seems to be a resounding yes. That is why China brings them up at every meeting and asks for them to be relaxed. It's really hurting them domestically - they have to rely on smuggled SOTA chips for any meaningful advancements while they wait for their domestic capabilities to ramp up (which will take years).
I don't have the answer, but I can understand the viewpoint that China's temporary kneecapping may actually lead to long term supremacy, as their in-country solutions become capable of competing with the state of the art. That will leave America more vulnerable in relation to China, because we will still be relying on access to technology from a wide range of countries (Netherlands, Taiwan, South Korea) in order to compete. That gives China additional leverage over the United States, as we will remain reliant on international cooperation.
And this analysis doesn't even address the ramifications of China exporting this technology, increasing their export dominance and potentially overtaking America's tech dominance at the software and design level of the stack.
I don't know what the right answer is to the problem, but it doesn't take much effort to imagine our current efforts as being the wrong answer, which is a little troubling.
It's hard to say that any US AI companies have benefited from these sanctions (or that this was the goal). Who has benefited, and how? US companies are still having to compete with them, with no restriction on Chinese AI being used in the US (e.g. GLM available via Amazon Bedrock, DeepInfra, Fireworks AI, etc). NVIDIA is losing sales and CUDA lock-in. Other supply chain vendors losing sales too.
I mean, I remember listening to the Biden people back in 2022 talking how they were going to cripple China's semis and therefore AI industry and keep them 5+ years behind the curve as Team America accelerates ahead. That was the pitch.
You've now got Huawei Ascend 950, GLM-5.2 at Opus 4.8 levels, China dominating OSS models, and Z.ai saying they'll have a Fable-level model by EOY. I would say the export controls have utterly, utterly failed.
https://www.csis.org/analysis/understanding-biden-administra...
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.
The vast, vast majority of websites have no business of using so much JS, especially in critical paths, anyways.