China's AI Companies May Be 'Distilling' America's AI Models (yahoo.com) 51
In March, Anthropic's Claude "quietly deployed software to spy on China-based customers," reports the Washington Post — apparently to unmask Chinese rivals "suspected of hijacking its technology to make their own AI tools smarter."
Last week Anthropic removed the spyware "after a software developer revealed its existence and privacy advocates criticized Anthropic, saying it had surveilled its own users."
Anthropic's tracking code was designed in part to catch Chinese firms "distilling" its AI models, a technique that involves pressing a large, expensive AI system to serve as a tutor to a smaller, cheaper one. Asking the larger system huge numbers of questions — hundreds of thousands or more — generates responses that can be used to upgrade the power of the smaller one on the cheap. Distillation isn't illegal, and it has been used for years in the AI industry. But distillation without permission is against AI companies' rules, and, used effectively, is giving Chinese AI companies a major leg up, American AI companies say... Anthropic and ChatGPT-maker OpenAI have both accused Chinese AI companies of using this technique to build copycat AI models of their own.
In a May blog post, Anthropic said that Chinese companies' use of distillation, along with evading U.S. export controls on high-end computer chips, has allowed them to "trail closely" behind U.S. models. But if these techniques can be blocked, it might be possible for the United States to "lock in a 12-24 month lead" on Chinese capabilities, the company said... This month, Anthropic said in a letter to U.S. senators that was obtained by The Post that it uncovered a campaign in which Chinese tech giant Alibaba's Qwen AI team used roughly 25,000 fraudulent accounts to generate more than 28.8 million exchanges with Claude to improve its own technology. In February, Anthropic made similar accusations against the Chinese firms Deepseek, Moonshot and MiniMax and said the campaigns were "growing in intensity and sophistication...." Anthropic and OpenAI have appealed to the U.S. government, arguing that distillation amounts to intellectual property theft that harms the U.S. in the geopolitical AI contest....
That Chinese AI labs are using U.S. models to improve their own technology appears beyond dispute. In a February 2025 study, researchers from China's Peking University and the state-funded Chinese Academy of Sciences developed methods to detect signs of distillation in leading large language models. They concluded that, with the exception of ByteDance's Doubao, most domestic models they tested showed substantial evidence of distillation, mostly drawing from U.S. models... In one set of intensive tests, a Qwen model misidentified itself as Claude nearly a third of the time, the Chinese researchers found.
U.S. firms have also used distillation to piggyback on AI systems made by others. In 2024, OpenAI released a tool to make it easier for customers to distill its own models and produce data sets for AI training. SpaceX founder Elon Musk said in court testimony in May that his AI company xAI used distillation to train its models and that the technique is common throughout the industry.
The article also notes that Anthropic "said it has banned nearly 700,000 accounts that were using Claude in China." But the article includes this quote from Kyle Chan, a fellow at the Washington-based Brookings Institution's China Center. "Anthropic's framing is that this is a geopolitical contest for basically the future of the world and freedom and democracy. It's that this is not just undercutting the U.S. commercially, but undercutting American strategic advantage in the most powerful technology we know today."
In a May blog post, Anthropic said that Chinese companies' use of distillation, along with evading U.S. export controls on high-end computer chips, has allowed them to "trail closely" behind U.S. models. But if these techniques can be blocked, it might be possible for the United States to "lock in a 12-24 month lead" on Chinese capabilities, the company said... This month, Anthropic said in a letter to U.S. senators that was obtained by The Post that it uncovered a campaign in which Chinese tech giant Alibaba's Qwen AI team used roughly 25,000 fraudulent accounts to generate more than 28.8 million exchanges with Claude to improve its own technology. In February, Anthropic made similar accusations against the Chinese firms Deepseek, Moonshot and MiniMax and said the campaigns were "growing in intensity and sophistication...." Anthropic and OpenAI have appealed to the U.S. government, arguing that distillation amounts to intellectual property theft that harms the U.S. in the geopolitical AI contest....
That Chinese AI labs are using U.S. models to improve their own technology appears beyond dispute. In a February 2025 study, researchers from China's Peking University and the state-funded Chinese Academy of Sciences developed methods to detect signs of distillation in leading large language models. They concluded that, with the exception of ByteDance's Doubao, most domestic models they tested showed substantial evidence of distillation, mostly drawing from U.S. models... In one set of intensive tests, a Qwen model misidentified itself as Claude nearly a third of the time, the Chinese researchers found.
U.S. firms have also used distillation to piggyback on AI systems made by others. In 2024, OpenAI released a tool to make it easier for customers to distill its own models and produce data sets for AI training. SpaceX founder Elon Musk said in court testimony in May that his AI company xAI used distillation to train its models and that the technique is common throughout the industry.
The article also notes that Anthropic "said it has banned nearly 700,000 accounts that were using Claude in China." But the article includes this quote from Kyle Chan, a fellow at the Washington-based Brookings Institution's China Center. "Anthropic's framing is that this is a geopolitical contest for basically the future of the world and freedom and democracy. It's that this is not just undercutting the U.S. commercially, but undercutting American strategic advantage in the most powerful technology we know today."
And? (Score:5, Insightful)
American AI companies 'distilled' millions of works from the original authors, they dont like it? Tough.
I don't care where things are made anymore (Score:3, Insightful)
It's 2026 is there anyone naive enough to believe that there's some magical world where jobs come back? Half of people under 30 are living with their parents and no it's not because of cell phones and avocado toast...
Re: (Score:2)
It's okay if you don't reply to your post as ACoward, and then reply to ACoward's post as ACoward as if the ACs are two separate entities
Re: (Score:3)
Well quite. Either making an AI by copying something is copying in which case they're a bunch of criminals ready to go to prison for a thousand years according to how it works for peons, or it's not in which case the Chinese are also doing nothing wrong.
They want it both ways,
Utter fuckers.
And? (Score:5, Insightful)
Re: (Score:2)
And Anthropic created content that is not copyrighted and cries foul when someone uses it like it were not copyrighted. It is even unclear if they could really ban you for that, but of course they can ban using their "... or without a reason" clauses.
China and everyone else (Score:3)
You make a question answering web service and then people pay you to have it answer their questions. How unfair. Oh the humanity.
Re: (Score:1)
The llms aren't a 'question answering service'. They are a pattern-matching and string generating service.
These systems cannot 'answer questions' because they don't understand them.
A simple test: ask the same question in a positive and a negative tone and watch the "answering service" generate two completely different answers, simply because the generator is producing output based on different matches.
Re: (Score:1)
"These systems cannot 'answer questions' because they don't understand them."
This is obviously false. AI systems are already answering questions, and whether or not the AI "understands" it is irrelevant if the information being returned is largely correct.
Sure, there may sometimes be some incorrect information or hallucinations, but most questions on general subjects do seem to yield reasonable results, and it's only when you get into areas that have small volumes of training data that you tend to get sketc
Re: (Score:2)
Blah blah blah assumption, blah. ACs are copy/paste services.
Anthropic is using... (Score:5, Insightful)
...every dirty trick in the book to secure their monopoly.
They spread fear and claim that only they can ensure safety.
They refuse to release a powerful model, claiming it's too dangerous.
They invite government restrictions.
They attack open source projects.
Classic monopolist behavior.
Re: Anthropic is using... (Score:3)
Re: (Score:3)
It depends on what you’re doing. Claude is the best for generating code.
Monopoly is inevitable (Score:2, Insightful)
That means only the big platform holders are going to be able to keep their models fed and current.
So facebook, microsoft, maybe Apple and that's about it. In the past I would include Twitter but I think it's mostly politics bots and pussy pics in bio posts.
AI is a technology that by its very nature becomes a monopoly.
Re:Monopoly is inevitable (Score:5, Interesting)
Do they retain all their training data? can they store all that? - i thought they were using all the internet and massive piracy?
The web is being polluted with slop so.... I would think China could get around all copy-protection and have an advantage in data collection outside of the slop invested parts of the web. If the USA AI corps were not violating the law, they'd be trying to scrape from China's bots who don't have their legal limitations... Is the old web even that valuable to mine in the 1st place? From what I've read these AI are pretty amazing when reduced to a relatively small domain data set; like all journals and books on 1 topic. Have you not tried to research something in depth on the web and found it to be severely lacking compared to books and journals?? Even online lectures are just highlights from textbooks... Well, when you look to the book... I've spent years reading on the web on a topic for entertainment then tried a book only to find it had everything I learned all in 1 concise place that would have taken a fraction of the time and effort... and without all the filtering and correcting of know-it-all blowhard slop and that was before we automated windbags with AI.
Tech makes things worse. It's like a drug. Opioid... It has targeted controlled good use cases but outside of that it's bad stuff. Everybody's answer to the problems it creates is to get more tech...at at least suffer until the next update/version... People were already getting more stupid, especially in the USA now we have AI and already we have studies showing it does just that... If you think things are stupid now...
With a 15% drop in PhDs ,science going down hill , and CS people leaving or souring on the evil of the corps here... I think these tech bros are quite self inflated as to their importance and how close they are to the end game. They are not going to get their huge break-thru monopoly they are racing towards like mad which looks more like a cover for a Ponzi scheme hoping to become a real business before it collapses. 80% of the effort is for the last 20%. They are probably not even at the last 20% and even then, their "AGI" could take centuries to get the last 5% of it (if you can even measure it well enough know when you are at 95%... or even at 80% progress. Assuming, you know what 100% even is!)
Re: (Score:3)
You assume that there is value in "keeping your models current", which is a questionable proposition.
I have a couple of (rather simple) systems that successfully use "non-current" models that access a database with current data as needed.
The setup works quite well for code generation that must conform to APIs that the original model has never seen - in fact, not much worse than the most "frontier" among the "frontier models".
Re: (Score:1)
And as usual they claim that China is only capable of stealing and copying. It's always the same mistake, and the same outcome.
Re: (Score:2)
The main problem they have is that their monopoly is not actually one. That it is all made up.
Re: (Score:2)
Synthetic data contains too many model names to prevent that from happening. Even without direct distilling you can bet that they incorporate every dataset they can find in the different repositories. And many of that are distilled from a wealth of different models. Anthropic would be stupid if they wouldn't do that themselves as well.
Completely normal and expected. (Score:2)
I understand! (Score:2, Insightful)
Theft of IP is only OK when large American companies do it.
When I was young, I thought people blathering on about class war were propagandized idiots. Turns out I was the propagandized one.
People generally act based on their own selfish interests, and the rich want to be richer. They can buy policy, we can't. They are insulated from us by their wealth and we don't matter. We have no rights, we're not people because we're not rich. They can steal from us but can then wield the power of the government to
Can someone explain how Anthropic was spying? (Score:1)
Oh No! (Score:3)
Anyway. Spies gonna spy. Counterintelligence gonna spy on the spies. (Yo dawg...) If you don't like it get out of the game.
These companies are constantly one-upping each other. Tight competition means they are trying anything for an edge.
Let 'em play, ref!
Does a bear d something in the woods? (Score:3)
Yeah, same joke but AIs don't call for original humor. Still, I sadly thought the story had more potential for funny.
Fixed that for you (Score:2, Funny)
Anthropic's framing is that this is a geopolitical contest for basically the future of the world and freedom and democracy; assuming in Nov Trump is voted out and Democratic Socialist voted in.
No go! (Score:1)
Don't steal what we have stolen.
I don't mind everyone distilling everyone. If this means better models in the end, it's fine. And the Chinese companies release most their models, so it's only about OpenAI and Anthropic if they indeed distill. I would also think they might be crying wolf all the time so daddy Trump bans access to Chinese services and open models.
Distill this, Xi! (Score:1)
Tank Man Tienanmen Pooh Bear Country of Taiwan Tank Man Tienanmen Pooh Bear Tank Man Tienanmen Pooh Bear Country of Taiwan Tank Man Tienanmen Pooh Bear
Re: (Score:2)
Because the website helpfully provides structured data with tags which make it clear what it is, it's very easy to just use old fashioned and cheap filters to exclude content like that from the training set without losing anything else.
Nothing new (Score:2)
Deepseek originally identified itself as ChatGPT. It's pretty well known that the early versions of Deepseek were trained off one or more of the GPT4 models.
Until these Chinese state-backed firms start training their own models, they'll always be playing second fiddle.
Oh no! (Score:2)
'Distilling' America's AI Models (Score:2)
Unlikely (Score:2)
The amount of queries needed to steal a model (which is possible though queries) is prohibitive. This is probably just an attempt to keep up the fantasy that China cannot do their own models. That is nonsense. For example, small Switzerland has created its own open model (https://apertus-ai.org/), mostly as an academic project. China for sure can do it. All I see China doing is comparative benchmarks and such and that is perfectly legitimate.
If China can do it, that means the likes of Anthropic are not near
Building silicon gods (Score:2)
Most people noticed the demand to stop China using the same cheating and piracy the USA uses to build its machines.
Did you notice the "powerful technology" propaganda?
Powerful at what? Anthropic and OpenAI are still painting statistical word generators as gods in training. In truth, these gods, like 5 year-olds, can't be left alone for any useful time. The industry still hopes that throwing more and more CPUs/memory at the problem means, that one day, their gods will be infallible.
USA... What a shame!!! (Score:1)
It so ridiculous, so hypocrite...
When will people start to understand that "competition" is a prehistorical strategy that should be replaced by co-operation, should they wish to be recognised as "human beings"?