The U.S. says China's AI progress is down to 'distillation.' But is it that clear cut?

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

In the battle for AI supremacy between the U.S. and China, "distillation" has become a hot topic.

The term broadly refers to the training of an AI model using the output or answers from another more advanced model. In some instances the process has been deemed controversial because a less advanced lab could use the product of a model that has taken millions or billions of dollars to create to design a competitive product.

As China has continued to pump out highly advanced models which compete in some areas with their American rivals, U.S. companies and Washington have accused Chinese labs of using distillation to gain ground, denouncing their efforts as theft.

But Aidan Gomez, the CEO of Cohere, says not all of China's progress is down to this technique.

Gomez, who spoke to me on the latest episode of The Tech Download, is an authoritative figure in AI. He was a co-author of a 2017 research paper called "Attention Is All You Need," which underpins many of the AI services we use today such as ChatGPT and Claude.

Gomez described Chinese AI models as "world class," adding that the lead U.S. labs have is "evaporating very quickly."

"There was like a lot of talk about the Chinese are just copying, they're just distilling, they're cheating. A lot of these tropes were applied, and that was definitely true to an extent," Gomez told me.

"However, they have developed an exceptional capability independent of distillation, and it's proven by the fact that the latest models that are coming out, actually on some benchmarks, on some axis capabilities, beat the best American models. And you can't copy or distill to better. You can close the gap and reduce the gap by copying, but you can't outperform."

This runs slightly counter to what has been said by U.S. AI labs and Washington.

Before I get to that however, I think it's important to add some context about the development of the Chinese tech industry which I witnessed first hand while reporting from the country for three years. 

The talent pool is huge, the pace of innovation is rapid and when an industry is backed by the central government in the way that it has been, it certainly adds a strong tailwind. China has often been labelled with the "copycat" moniker but in reality, that has faded as real innovation has come out of the tech scene there over the last few years from electric cars, to robotics and even advancements on semiconductors.

Gomez said there is some distillation taking place but that can't explain all of China's progress on AI models. That view runs contrary to some of the warnings being sounded by U.S. AI labs.

Anthropic, the creator of Claude, has on numerous occasions said Chinese AI labs have used distillation, saying it amounted to theft. This month, Anthropic's head of threat intelligence, Jacob Klein, told CNBC that "there's an entire illicit ecosystem to try to gain access to Claude and other models."

In a report released this month, Anthropic said companies including Alibaba, Moonshot and DeepSeek, which are among China's leading AI labs, attempted to use "illicit distillation" to train their own models.

The U.S. government has also warned about distillation on multiple occasions. Earlier this month, the Cybersecurity and Infrastructure Security Agency said Chinese AI companies are "conducting systematic extraction of proprietary functionalities and capabilities of U.S. AI companies' models through industrial-scale knowledge distillation campaigns that form the core—not merely a supplement—of their AI development strategy."

However, there is a growing chorus of voices joining Gomez in casting doubt on how important distillation is to Chinese AI model development.

Sriram Krishnan, a former senior White House policy advisor on artificial intelligence, said Thursday that services like ChatGPT and Claude "came out of distilling human content," referring to the fact that models for these products were trained on information on the internet.

"So the idea of distilling has always been a core part of how computer science works," Krishnan told CNBC's "Squawk Box." 

According to Krishnan, even if Chinese labs are using distillation, it's unclear how much this is actually adding to their ability to advance models, a similar point of view to Cohere's Gomez.

The Chinese government has hit back at distillation claims. A China Ministry of Commerce spokesperson said the U.S.'s allegations that Chinese companies are using "industrial-scale" distillation is "groundless and legally unsound."

China's progress in AI has come even as the country faces various restrictions on access to cutting-edge chips from companies like Nvidia. Instead, China has ramped up the development of its homegrown semiconductor sector. Even though most experts agree that China's semiconductors do not compete with the performance of U.S. chips, China has used various techniques to make them work.

Analysts have highlighted the lack of access to key technology as a reason for Chinese innovation.

"Because of chip restrictions, Chinese developers were forced to build leaner, smarter architectures and do more with less compute," Neil Shah, a partner at Counterpoint Research, told CNBC. "Dismissing that as mere imitation might make for convenient restrictive policy making, but it fundamentally misjudges the competition," Shah added.

It raises the question: Are Western policymakers underestimating the technical capabilities of China?

"Absolutely," Cohere's Gomez told me. "It takes a while for people's mindset to change and to recognize a shift. We're still reckoning with the fact that China is no longer just copying but is actually innovating outright."

Anthropic and OpenAI are hunting for smaller AI data center deals, sources told CNBC, as the race to access the infrastructure needed to deploy workloads ramps up.

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

The U.S. now faces a shortage of up to 157,000 semiconductor workers by 2030, according to a new report.

The Chinese AI startup behind the Kimi model said Thursday that financial industry giants, including investment banks, funds and venture capital firms are using its models on AI tools.

AWS said Tuesday it can't restore access to cloud-computing facilities in Bahrain and parts of the United Arab Emirates, more than six months after the sites were damaged by drone strikes in the early days of the Iran war.

As researchers warn that AI could pose catastrophic risks, some of the pioneers who helped develop the technology, including Yoshua Bengio and Geoffrey Hinton, have weighed in with their own warnings.