From Silicon Valley to DC, the tech world is full of AI

Earlier this year, Google AI leader Jeff Dean, in a podcast, discussed a concept that, at the time, was not talked about outside of wonky tech circles: distillation.

When talking about the development of Google’s AI models, Dean said that he and his colleagues discovered techniques for the distillation of intelligence because Google was looking to improve the performance of its systems without relying on one large image recognition model.

“With distillation, which is the main way to make small models work, you have to have a boundary model to put into your small model,” Dean said in February.

Five months later, distillation has suddenly become a hot topic from Silicon Valley to Washington, DC, as techies and lawmakers argue that the practice is becoming a national security threat and allows China to catch up with the US in the high-level AI race. Concerns arose late last week after Chinese lab Moonshot AI released the Kimi K3, and users quickly found it competing with more commercially available AI from Anthropic and OpenAI.

Unlike leading US AI companies, which sell access to proprietary models, Moonshot and other Chinese labs offer so-called open-source models that allow users to download the technology, modify it and use it wherever they want.

Some government officials say that Moonshot’s ability to be found so quickly in the distillation, describes it as theft of American intellectual property, especially by incorporating the Anthropic Fable model of the frontier.

“We have information that Moonshot AI released Anthropic’s Fable to improve its K3 model,” White House adviser Michael Kratsios told X on Wednesday. “To do this they have developed a sophisticated internal platform to conduct distillation on a large scale against US models, allowing them to quickly switch between multiple access methods to avoid detection.”

At a high level, distillation refers to the use of responses from a chatbot or work product from an advanced AI model to train another model. This practice is controversial because, depending on how it is used, it can allow a model developer to make a competitive offering simply by using the output from companies that have invested millions or billions of dollars developing highly sophisticated training technologies.

“It’s like someone went to a course, read a book, and did all the hard work of doing homework,” said Pukar Hamal, founder of AI security firm SecurityPal. “Then another student is like, ‘Hey, I didn’t do that. Can I copy your work?’

Whether it was Kratsios’ post or something else, the world’s biggest tech heavyweights came together on Friday in an unprecedented way to make their position clear. After a series of social media posts throughout the week, tech giants Nvidia, Microsoft, Meta, Palantir joined more than 20 other companies to issue a letter urging policymakers to avoid “premature restrictions” on open-source AI models that “could stifle competition or drive innovation overseas.”

“Distillation, or the practice of using the results of one model to help train or improve another, is a widely used technique for model development, evolution, and validation,” they write.

The emergence of distillation presents a conundrum for US policymakers, who have long worried about Chinese technology over IP theft and national security issues.

Colin Shea-Blymyer, a researcher at Georgetown’s Center for Security and Emerging Technology, said the US government is trying to figure out where it stands.

The government could argue that Chinese and Russian companies “have used the results of the American labor models to make them more efficient, so they have an unfair advantage there,” Shea-Blymyer said.

The box CEO Aaron Levie was one of the signatories of Friday’s letter. Levie said in an interview that in order to remain competitive, American companies must be able to access high-quality technology, regardless of where it was developed.

“Generally the arc will be that the innovation that’s there, whether it’s from the US or China or otherwise, you have to wait for AI progress, and it’s usually going to be more cost-effective and more efficient in the long run,” Levie said.

While most of the talk now focuses on China’s open-weight AI models such as the Kimi K3, many companies have included a milling method when making their models, says Shashi Bellamkonda, director of research at Info-Tech Research Group. Nvidia, for example, used filtering as part of the training process in a series of Llama Nemotron models, as explained in the accompanying research paper.

“It’s a legitimate and very important way to train a small, cheap model from the output of a large model, and it’s done all the time,” Bellamkonda said.

However, Anthropic has a different perspective, because the company sees how its models are used and has a growing business to protect. In February, the company said its Claude power was breached on an “industry scale” by China’s DeepSeek, Moonshot, and MiniMax, which used 24,000 fake accounts, generating 16 million transactions.

Anthropic, which is valued at about $1 trillion and has ambitions to go public soon, has said that stopping illegal drilling is a matter of national security.

“Anthropic and other US companies are developing systems that prevent state and non-state actors from using AI to, for example, develop bioweapons or conduct malicious cyber operations,” the company said in its February post. And stopping it requires “immediate, coordinated action among industry players, policymakers, and the global AI community.”

OpenAI and Anthropic prohibit distilling in their terms of service. Bellamkonda said they are actually suggesting that using their super models without authorization represents potential IP theft.

But with the cost of AI rising, companies will do whatever it takes to be successful.

Hamal said he would have no problem using lightweight Chinese models like the Kimi K3 from SecurityPal, which performs security checks using AI. He says it can save them a lot of money.

“We will make sure that there are no bad backends in this code,” Hamal said. “But to handle the infrastructure ourselves after doing the test, why not?”

One big problem for Anthropic and OpenAI as they try to make their case about IP theft is that both companies rely on other sources of content to build their models, and have been sued for doing so.

Max Pritt, an attorney at Boies Schiller Flexner who represents authors in patent cases against AI firms, said the government is in the same boat.

“Administrations, at least publicly, have focused their efforts on protecting the intellectual property of technology companies, while presenting a large part about the property of creators and individual property that was used without authorization,” Pritt said.

WATCH: China’s AI companies are finding ways to make money as their models remain open