Has China’s AI progress really reached the point of “distillation”?

This report is from this week’s The Tech Download newsletter. Is it what you see? You can subscribe here.

Distillation has become a hot topic in the battle for AI supremacy between the US and China.

This term broadly refers to training an AI model using output or answers from another, more advanced model. It is believed that this process can be controversial because in some cases, less advanced labs may use models that cost millions or billions of dollars to create to design competitive products.

As China continues to roll out advanced models that compete with U.S. rivals in some areas, U.S. companies and Washington have accused Chinese labs of exploiting distillation for profit and denounced the effort as theft.

But Kohia CEO Aidan Gomez says not all of China’s progress is due to this technology.

Gomez, who spoke to me on the latest episode of The Tech Download, is an authority on AI. He is a co-author of the 2017 research paper “Attending Is All You Need,” which powers many of the AI ​​services in use today, including ChatGPT and Claude.

Gomez described China’s AI models as “world class,” adding that the lead held by U.S. labs is “quickly evaporating.”

“There was a lot of talk that the Chinese were just copying, that they were just distilling, that they were cheating. A lot of these metaphors were applied, and it was certainly true to a certain extent,” Gomez told me.

“But they’ve developed extraordinary capabilities that are independent of distillation, and that’s evidenced by the fact that the latest models that are coming out actually outperform the best American models on some benchmarks and some axis capabilities. And you can’t copy or distill to make it better. You can close the gap by copying, you can reduce the gap, but you can’t outperform.”

This is somewhat contrary to what the US AI Institute and the Washington government have said.

But before we get into that, I think it’s important to add some background on the development of China’s technology industry, which I witnessed firsthand during my three years reporting from China.

The talent pool is huge, the pace of innovation is fast, and if the industry is supported by central government, as it has been in the past, it will certainly add a strong tailwind. China has often been labeled with the nickname “copycat,” but that image has actually faded in recent years as real innovations have emerged from the country’s tech scene, from electric cars to robotics to advances in semiconductors.

Gomez said there is some distillation going on, but it doesn’t explain all of China’s progress in AI models. This view contradicts some of the warnings issued by the US AI Institute.

Claude’s developer, Anthropic, has repeatedly said that Chinese AI labs are using distillation, saying it amounts to theft. “There’s a whole illicit ecosystem out there trying to access Claude and other models,” Jacob Klein, head of threat intelligence at Anthropic, told CNBC this month.

Anthropic said in a report released this month that companies including Alibaba, one of China’s leading AI labs, Moonshot and DeepSeek have attempted to use “illegal distillation” to train their models.

The US government has also warned many times about distillation. The Cybersecurity and Infrastructure Security Agency said earlier this month that Chinese AI companies are “systematically distilling the unique capabilities and capabilities of the U.S. AI enterprise model through industrial-scale knowledge distillation campaigns that are core to, rather than merely supplementary to, AI development strategies.”

But Gomez is joined by a growing number of voices questioning the importance of distillation to China’s AI model development.

Sriram Krishnan, a former White House senior policy adviser on artificial intelligence, said Thursday that services like ChatGPT and Claude were “born out of extracting human content,” referring to the fact that the models in these products were trained based on information on the Internet.

“So the idea of ​​distillation has always been at the core of how computer science works,” Krishnan said on CNBC’s “Squawk Box.”

Krishnan says that even if Chinese labs are using distillation, it’s unclear how much it actually adds to their ability to advance the model, a view echoed by Kohia’s Gomez.

The Chinese government has pushed back against the distillation claims. A spokesperson for China’s Ministry of Commerce said US claims that Chinese companies were using “industrial-scale” distillation were “baseless and legally unwarranted.”

China is making advances in AI despite facing various restrictions on access to cutting-edge chips for companies such as: Nvidia. Instead, China is ramping up development of its homegrown semiconductor sector. Even though most experts agree that Chinese semiconductors do not compete with the performance of American chips, China uses different technologies to make its semiconductors work.

Analysts highlight China’s lack of access to key technologies as a reason for its innovation.

“Chip limitations have forced Chinese developers to build leaner, smarter architectures and do more with less computing,” Neil Shah, partner at Counterpoint Research, told CNBC. “Dismissing it as mere imitation may enable convenient and restrictive policy decisions, but it fundamentally misjudges competition,” Shah added.

The question arises whether Western policymakers are underestimating China’s technological capabilities.

“That’s right,” Kohia’s Gomez told me. “It takes time for people’s mindsets to change and for them to recognize change. We are still considering the fact that China is no longer just copying, but is actually completely innovating.”

Anthropic and OpenAI are exploring smaller AI data center deals as competition for access to the infrastructure needed to deploy workloads intensifies, people told CNBC.

Nvidia will double the number of chips it sells next year, CEO Jensen Huang said Thursday.

The United States currently faces a shortage of up to 157,000 semiconductor workers by 2030, according to a new report.

The Chinese AI startup behind Kimi Model announced Thursday that giants in the financial industry, including investment banks, funds and venture capital firms, are using its models in their AI tools.

AWS Access to cloud computing facilities in parts of Bahrain and the United Arab Emirates cannot be restored, more than six months after they were damaged in a drone strike early in the Iran war, it said on Tuesday.

As researchers warn that AI could pose devastating risks, some of the pioneers who helped develop the technology, including Joshua Bengio and Jeffrey Hinton, are speaking out with their own warnings.