China’s DeepSeek
Top White House advisers have raised alarms over China’s DeepSeek, a tech company that has reportedly benefited from a method known as “distillation,” which allegedly allows it to capitalize on the advancements of U.S. AI models.
This technique, wherein one AI system learns from another, could be difficult to block, according to sources in Silicon Valley. The controversy erupted after DeepSeek released a new AI model that rivals the capabilities of U.S. tech giants like OpenAI, offering it at a significantly lower cost and making the code publicly available.
Distillation involves having a newer, smaller AI model learn from an older, more powerful one by evaluating the quality of its responses. This allows the newer model to reap the benefits of the considerable time and resources invested in the older model, without incurring similar expenses.
While distillation is a common practice in AI development, some view DeepSeek’s use of this method as a violation of the terms of service set by prominent U.S. companies, including OpenAI.
OpenAI has confirmed that it is aware of efforts in China to replicate U.S. AI models using distillation and is reviewing whether DeepSeek has misused its technology.
Naveen Rao, vice president of AI at Databricks, emphasized that learning from competitors is a common practice in the AI industry, much like how automakers analyze each other’s engines to improve their own.
However, U.S. officials have voiced concerns, with Howard Lutnick, President Donald Trump’s nominee for Secretary of Commerce, accusing DeepSeek of misappropriating U.S. AI technology and promising to impose export controls. White House AI czar David Sacks also expressed his concerns on the matter.
Technologists have noted that blocking distillation is not as straightforward as it may seem. DeepSeek’s innovation includes showing that even a small number of data samples from a larger, more capable model can dramatically enhance the performance of a smaller model.
Since AI products like ChatGPT have hundreds of millions of users, detecting such violations could be challenging, especially since some models, like Meta’s Llama and French startup Mistral’s offerings, are open-source and can be freely downloaded for use in private data centers.
Umesh Padval, managing director at Thomvest Ventures, pointed out that stopping distillation from open-source models such as Mistral and Llama is nearly impossible, as these models are readily accessible to anyone.
While Meta requires users of its Llama model to disclose distillation practices, DeepSeek has only acknowledged using Llama for some of its models but did not address whether it used Meta’s model in earlier versions.
Blocking this practice entirely may require stringent “know-your-customer” policies, similar to those used in the financial sector, but such measures are not yet in place.
Jonathan Ross, CEO of AI computing company Groq, has taken the step of blocking all Chinese IP addresses from accessing his cloud, in an attempt to prevent Chinese companies from exploiting AI models hosted there.
However, Ross acknowledged that this strategy is insufficient and warned that it would become a “cat and mouse game,” as people will inevitably find ways around these measures. The challenge remains: finding a comprehensive solution to prevent firms like DeepSeek from taking advantage of U.S. AI models.
