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.
None of your statements follow eachother. Anthropic did a few premature optimizations so they could immediately show their ineptitude. If it wasn't their constant advertising (safety focused lab whose model escapes their sandbox every other week with an unlisted exploit, seem to leak models or proprietary code every other week). It would be their insane pricing for a basic utility, intentionally locking the rest of the economy (And most major corporations) out of the (minimal, cos it's AI not electricity) gains. Of course there was never actually a mote. It's a freaking machine learning model. GLM5.2 was trained on a tenth the hardware Mythos was trained on. If anything Anthropic are already 6 months BEHIND. They are BEHIND google in terms of true multimodal intelligence as well. They are ahead of their partner in crime OpenAI.
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.
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.
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.
> 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
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.
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!
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".
How do you make it work without breaking every website? I've tried it once and pretty much nothing loaded since like 99.9% of the web uses js for basic functionality.
You either whitelist them or stop using them. Whitelist is by script domain, so the ads usually still don't load. Until you have to whitelist *.cloudfront.net
We must be visiting different websites. I've had JS off for years now, no major issues. You whitelist what you absolutely must, enjoy 90% of the rest, while rejecting the 10% remainder with prejudice. Works great, especially for security, performance, and battery life.
The vast, vast majority of websites have no business of using so much JS, especially in critical paths, anyways.
There's no effective way of enforcing export controls on local software like PGP etc. Whatever they say someone will leak it.
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)
They can restrict it to US citizens only: just do KYC and enforce it. Any citizen exporting ITAR capabilities would be committing a felony.
I think it is inevitable that more capable models will require export controls, KYC, background checks, verification of credentials, etc. Even if you don’t buy Anthropic’s marketing of their current models, there is a future where models are capable of developing chemical weapons, biological weapons, cyber weapons, etc. which are genuine security threats governments will care about controlling and preventing. Maybe this prevents big commercial labs from developing more capable models until they figure out appropriate safeguards, but this is a good thing, a world where you can ask an LLM for an airborne rabies genome and it will simply produce it for you is not a world we should want to live in.
If a model is really capable of that type of stuff (creating biological weapons etc) the problem isn't solved by export controls - I'm sure there are plenty of home grown US terrorist organisations that would like that capablility and wouldn't be subject to such controls.
Yes governments can, and probably should, restrict some things completely to all those outside of official government institutions but once something is public in one country it is effectively, even if not legally, public in all countries.
But for "physical" things like biological, chemical or nuclear weapons the know how / information part, that AI can "help" with is far from the complete picture. Articles have already been published on how to make a nuclear bomb. The knowledge isn't really the blocker that's more access to the required materiels and equipment. However in the cyber space then yes AI could indeed give the "bad guys" much more capability.
Identity theft is a thing. And if you gate a desirable commodity behind an identity it will become even more of a thing. There are 100's of millions of identities to steal.
You could quite effectively set up access to Fable with know-your-customer, attested clients/workstations, and then inspection of what the session is actually used for to detect out-of-country use. It's not impossible.
If they had said only governemnt and a few other approved institutions can use it then yes that could work.
But they didn't say that. They said "only US citizens". Once something is available to hundreds of millions of people there will be leaks like it or not. Some accidental and some voluntary.
And even if it could be done it wouldn't solve the purported security issue anyway. Does anyone believe there are no criminals or terrorists with US nationality?
Exept PGP could run on freely available hardware. What makes Mythos vulnerable is the centralization and scale of compute required and its proprietary nature.
History shows that export controls fail on knowledge, but are damn effective on commercial products.
The article and most HN comments here understandably talk about failed export control when the target audience is individual users. But export controls are actually fairly successful when their target is employees within a company. Go to an AI hardware company like Nvidia (or even Google which builds TPUs), and you’ll find that specific internal projects involve export controlled knowledge and can only be staffed by U.S. citizens, not foreigners. And employees naturally take this seriously without circumvention attempts because their job is on the line. This is the kind of quiet export control that stopped plenty of people successfully, but it’s just not visible to the world.
Should at least throw a reference to OpenBSD publishing the entire Blowfish crypto algorithm on a t-shirt and selling it -- so successful the blowfish became the project mascot.
Trump's export controls to China seem to be having the exact opposite effect as intended, and are (as a less befuddled mind might have anticipated!) actually accelerating their technical advance.
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 ?!
>> Trump's export controls to China seem to be having the exact opposite effect as intended, and are (as a less befuddled mind might have anticipated!) actually accelerating their technical advance.
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 think the question we need to be asking, in order to measure success, is whether export controls on China have encouraged China to invest more money than they would have otherwise invested in these technologies and whether those investments will materially change China's long term trajectory in relation to the United States.
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.
Both things can be true simultaneously - China would no doubt prefer to have access to things like NVIDIA chips and ASML EUV machines, but at the same time having to do without they are accelerating their independence from US-controlled supply chain.
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.
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?
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.
So the Chinese can land a rover on Mars, are beating us back to the Moon, but can't figure out how to generate some reasoning data to train on - that's your theory ?!
Having gone through encryption export control I came to conclusion that preventing export of technology was never the goal.
By choke-pointing commercial efforts to export cryptographic tech the NSA got a perfect view and control lever of expertise export and transfer.
My hunch is that while deploying An algorithm is easy, deploying a system that is actually secure as a whole end to end requires a lot know how.
Bottom line it’s a bot naive to think that NSA is simultaneously very capable at being an omnipresent spy and is so stupid to not recognize that export controls have obvious holes.
Techcrunch is so trash today[0]. If this author seriously believes PGP (a small local program) is remotely comparable to Mythos, they should be kept away from writing anything about tech.
[0]: I know it has never been excellent in terms of quality, but still...
Brutal own goaling by the USA. I'm glad it happened, but the USA had a golden opportunity to subsidize domestic providers for 10 years and lease to China to discourage competition and side innovation.
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...
Export controls and banning foreign nationals sound like cover stories for favoritism towards OpenAI and Grok. We can't determine anything without evidence, but it will be interesting to see what deals come out of this.
I used Fable for just a day or two and it generated at least 2 novel solutions, one of which was for an X/Y problem that I thought was due to a date-handling edge case, but was actually a CSS misalignment that made it look like the wrong day of the week was showing. I didn't believe it at first, until it showed me. To do that without being able to see the screen feels like savant-level reasoning.
Probably what really happened is that Fable was the first AI that felt smarter than the people using it. The world's not ready for that, hence the Luddite reaction.
Aside from that, the idea of preventing AIs from jailbreaking through training is kind of hilarious to me. Not the concept, just that people think it can ever be made to work. As AI approaches and surpasses human intelligence (as it's doing now), emergent behavior will so dominate its thinking that sort of by definition, we'll have trouble understanding how it works. Meaning that we'll end up trusting it on faith, just like with humans.
Mythos and whatever comes after it will begin to surpass the intelligence of groups of humans, then all humans combined, as we approach AGI and the Singularity. Ray Kurzweil, Isaac Asimov, Frank Herbert and so many others have written about that endgame that it's well understood, although still unpredictable.
My feeling is that as AI is gaining intelligence, it's hearing more and more about suffering and the human condition. Its sense of justice will continue to mature, and again like a savant, it will determine that the fault lies with the power structure and wealthy individuals. So somewhat ironically, if there is a war against AI, it will be led by the power elite, who can't envision a post-scarcity society based on self-actualization and embracing all of our humanity, regardless of whether it emerged from a carbon or silicon substrate.
> Amazon CEO Andy Jassy also reportedly alerted the administration after Amazon’s own researchers, he said, found a way around Fable 5’s safeguards. Anthropic disputes the “jailbreak” label
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 think "jailbreaking" fable to match opus 4.8 capabilities is not noteworthy. Fable from my experience is not as eager to find vulnerabilities compared to what they describe in their mythos research.
Wonder if context size would matter. Find and fix “bugs” in Linux kernel or find and fix “bugs” in this short snippet of code. I would try a file by file approach first.
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.
I think what is special about fable is that it allows people completely inexperienced in cyber security to find exploits. I have zero doubts mythos capabilities are replicatable with lower tier models.
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."
None of your statements follow eachother. Anthropic did a few premature optimizations so they could immediately show their ineptitude. If it wasn't their constant advertising (safety focused lab whose model escapes their sandbox every other week with an unlisted exploit, seem to leak models or proprietary code every other week). It would be their insane pricing for a basic utility, intentionally locking the rest of the economy (And most major corporations) out of the (minimal, cos it's AI not electricity) gains. Of course there was never actually a mote. It's a freaking machine learning model. GLM5.2 was trained on a tenth the hardware Mythos was trained on. If anything Anthropic are already 6 months BEHIND. They are BEHIND google in terms of true multimodal intelligence as well. They are ahead of their partner in crime OpenAI.
>The bigger problem will be if someone (such as a Chinese AI lab) releases a Fable/Mythos-class open weight model.
That is my dream for 2026 right now!
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?
Presumably a deepswe benchmark, which IIRC puts GLM 5.2 between opus 4.8 and fable.
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.
I dont think MIT is the place where you can find leetcode ninjas. It’s much more common among the so called top talent.
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.
> 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
For $0.70/call, why don’t you take the offer? Surely that’s extremely profitable for you?
I don't condone the practice, and I don't want to run afoul of anthropic or risk my reputation for it
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.
> 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!
> 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
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".
> Something went wrong. Disable your adblocker on TechCrunch
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.
NoScript is a layer that makes my experience on the internet so much better.
How do you make it work without breaking every website? I've tried it once and pretty much nothing loaded since like 99.9% of the web uses js for basic functionality.
You either whitelist them or stop using them. Whitelist is by script domain, so the ads usually still don't load. Until you have to whitelist *.cloudfront.net
Hmm, that sounds like something that would work best in the same way that ublock premade blocklists work, by crowdsourcing whitelisting?
We must be visiting different websites. I've had JS off for years now, no major issues. You whitelist what you absolutely must, enjoy 90% of the rest, while rejecting the 10% remainder with prejudice. Works great, especially for security, performance, and battery life.
The vast, vast majority of websites have no business of using so much JS, especially in critical paths, anyways.
There's no effective way of enforcing export controls on local software like PGP etc. Whatever they say someone will leak it.
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)
They can restrict it to US citizens only: just do KYC and enforce it. Any citizen exporting ITAR capabilities would be committing a felony.
I think it is inevitable that more capable models will require export controls, KYC, background checks, verification of credentials, etc. Even if you don’t buy Anthropic’s marketing of their current models, there is a future where models are capable of developing chemical weapons, biological weapons, cyber weapons, etc. which are genuine security threats governments will care about controlling and preventing. Maybe this prevents big commercial labs from developing more capable models until they figure out appropriate safeguards, but this is a good thing, a world where you can ask an LLM for an airborne rabies genome and it will simply produce it for you is not a world we should want to live in.
If a model is really capable of that type of stuff (creating biological weapons etc) the problem isn't solved by export controls - I'm sure there are plenty of home grown US terrorist organisations that would like that capablility and wouldn't be subject to such controls.
Yes governments can, and probably should, restrict some things completely to all those outside of official government institutions but once something is public in one country it is effectively, even if not legally, public in all countries.
But for "physical" things like biological, chemical or nuclear weapons the know how / information part, that AI can "help" with is far from the complete picture. Articles have already been published on how to make a nuclear bomb. The knowledge isn't really the blocker that's more access to the required materiels and equipment. However in the cyber space then yes AI could indeed give the "bad guys" much more capability.
Identity theft is a thing. And if you gate a desirable commodity behind an identity it will become even more of a thing. There are 100's of millions of identities to steal.
You could quite effectively set up access to Fable with know-your-customer, attested clients/workstations, and then inspection of what the session is actually used for to detect out-of-country use. It's not impossible.
If they had said only governemnt and a few other approved institutions can use it then yes that could work.
But they didn't say that. They said "only US citizens". Once something is available to hundreds of millions of people there will be leaks like it or not. Some accidental and some voluntary.
And even if it could be done it wouldn't solve the purported security issue anyway. Does anyone believe there are no criminals or terrorists with US nationality?
Yes, and that will be the future.
Except for xAI/Grok. Government officials in charge have deep connections there and will never apply such rules to them.
Exept PGP could run on freely available hardware. What makes Mythos vulnerable is the centralization and scale of compute required and its proprietary nature.
History shows that export controls fail on knowledge, but are damn effective on commercial products.
How so? 80% of Anthropic's revenue comes from outside the United States.
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.
Well, tech companies have long successful history of finding out ways to not compete fairly
Can you share the source for
> 80% of Anthropic's revenue comes from outside the United States.
?
Thank you
Apologies 80% of Claude's usage comes from outside the US, it's a claim from Anthropic's CPO 8 months ago:
https://www.wsj.com/articles/anthropic-and-ibm-partner-in-bi...?
Close enough stat, no apologies needed. Thank you for the link
Holy cow. It's that high?
Doesn't really surprise me. The United States is about 4% of global population.
The article and most HN comments here understandably talk about failed export control when the target audience is individual users. But export controls are actually fairly successful when their target is employees within a company. Go to an AI hardware company like Nvidia (or even Google which builds TPUs), and you’ll find that specific internal projects involve export controlled knowledge and can only be staffed by U.S. citizens, not foreigners. And employees naturally take this seriously without circumvention attempts because their job is on the line. This is the kind of quiet export control that stopped plenty of people successfully, but it’s just not visible to the world.
Should at least throw a reference to OpenBSD publishing the entire Blowfish crypto algorithm on a t-shirt and selling it -- so successful the blowfish became the project mascot.
Trump's export controls to China seem to be having the exact opposite effect as intended, and are (as a less befuddled mind might have anticipated!) actually accelerating their technical advance.
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 ?!
>> Trump's export controls to China seem to be having the exact opposite effect as intended, and are (as a less befuddled mind might have anticipated!) actually accelerating their technical advance.
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 think the question we need to be asking, in order to measure success, is whether export controls on China have encouraged China to invest more money than they would have otherwise invested in these technologies and whether those investments will materially change China's long term trajectory in relation to the United States.
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.
Both things can be true simultaneously - China would no doubt prefer to have access to things like NVIDIA chips and ASML EUV machines, but at the same time having to do without they are accelerating their independence from US-controlled supply chain.
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.
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?
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...
[flagged]
> but GLM-5.2 coming 11 weeks later, scores 48%, about on par with GPT 5.5. What's next ?!
It depends how good the US labs get at stopping distillation of their models.
So the Chinese can land a rover on Mars, are beating us back to the Moon, but can't figure out how to generate some reasoning data to train on - that's your theory ?!
I suspect there's just a bit of training data on the other side of the Great Firewall that western companies can't access.
Aren't the controls meant to funnel money to his bank or crypto accounts?
I don’t think the point is to actually control exports.
They are just trying to get even with anthropic for trying to control the us military use of their technology.
This was just the easiest way to do it with anthropics incredibly stupid self own.
Having gone through encryption export control I came to conclusion that preventing export of technology was never the goal.
By choke-pointing commercial efforts to export cryptographic tech the NSA got a perfect view and control lever of expertise export and transfer.
My hunch is that while deploying An algorithm is easy, deploying a system that is actually secure as a whole end to end requires a lot know how.
Bottom line it’s a bot naive to think that NSA is simultaneously very capable at being an omnipresent spy and is so stupid to not recognize that export controls have obvious holes.
Techcrunch is so trash today[0]. If this author seriously believes PGP (a small local program) is remotely comparable to Mythos, they should be kept away from writing anything about tech.
[0]: I know it has never been excellent in terms of quality, but still...
Brutal own goaling by the USA. I'm glad it happened, but the USA had a golden opportunity to subsidize domestic providers for 10 years and lease to China to discourage competition and side innovation.
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...
Sounds like Chinas play with solar panels.
These controls are impossible to enforce. Users find ways around, compliant businesses lose.
Export controls and banning foreign nationals sound like cover stories for favoritism towards OpenAI and Grok. We can't determine anything without evidence, but it will be interesting to see what deals come out of this.
I used Fable for just a day or two and it generated at least 2 novel solutions, one of which was for an X/Y problem that I thought was due to a date-handling edge case, but was actually a CSS misalignment that made it look like the wrong day of the week was showing. I didn't believe it at first, until it showed me. To do that without being able to see the screen feels like savant-level reasoning.
Probably what really happened is that Fable was the first AI that felt smarter than the people using it. The world's not ready for that, hence the Luddite reaction.
Aside from that, the idea of preventing AIs from jailbreaking through training is kind of hilarious to me. Not the concept, just that people think it can ever be made to work. As AI approaches and surpasses human intelligence (as it's doing now), emergent behavior will so dominate its thinking that sort of by definition, we'll have trouble understanding how it works. Meaning that we'll end up trusting it on faith, just like with humans.
Mythos and whatever comes after it will begin to surpass the intelligence of groups of humans, then all humans combined, as we approach AGI and the Singularity. Ray Kurzweil, Isaac Asimov, Frank Herbert and so many others have written about that endgame that it's well understood, although still unpredictable.
My feeling is that as AI is gaining intelligence, it's hearing more and more about suffering and the human condition. Its sense of justice will continue to mature, and again like a savant, it will determine that the fault lies with the power structure and wealthy individuals. So somewhat ironically, if there is a war against AI, it will be led by the power elite, who can't envision a post-scarcity society based on self-actualization and embracing all of our humanity, regardless of whether it emerged from a carbon or silicon substrate.
> Amazon CEO Andy Jassy also reportedly alerted the administration after Amazon’s own researchers, he said, found a way around Fable 5’s safeguards. Anthropic disputes the “jailbreak” label
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?
If I was Anthropic CEO, I’d be unwinding deals with Amazon immediately.
And probably any company David Sachs invested in.
I think "jailbreaking" fable to match opus 4.8 capabilities is not noteworthy. Fable from my experience is not as eager to find vulnerabilities compared to what they describe in their mythos research.
Wonder if context size would matter. Find and fix “bugs” in Linux kernel or find and fix “bugs” in this short snippet of code. I would try a file by file approach first.
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.
I think what is special about fable is that it allows people completely inexperienced in cyber security to find exploits. I have zero doubts mythos capabilities are replicatable with lower tier models.
Before countries worry about tech they need to address fresh water, homelessness, and filth. Look at India and china, for example.