Challenges in Blocking DeepSeek's Use of US Tech

This week, top advisers from the White House expressed concerns about China’s DeepSeek, a tech company believed to have benefited from a method referred to as "distillation." This technique allows one AI system to learn from another, making it challenging for authorities to stop such practices.

DeepSeek made headlines recently by releasing a new artificial intelligence model that claims to rival the capabilities of major US companies like OpenAI, but at a significantly reduced cost. In an unusual move, the company offered its code for free.

Experts in technology speculate that DeepSeek’s model has gained some of its advancements by analyzing American AI models. The distillation method encompasses an older and more robust AI model assessing the quality of responses generated by a newer model, thus transferring knowledge from the established system to the newer one. This allows the new model to utilize the vast investments made in developing the older model without incurring similar expenses.

This type of distillation is standard in the AI industry, although it violates the terms of service outlined by several leading US tech firms in recent years, including OpenAI. A spokesperson for OpenAI mentioned that they are aware of groups in China attempting to replicate US AI models through this method and are currently evaluating whether DeepSeek has improperly distilled their models.

Industry experts warned that it may prove difficult for the US government to prevent DeepSeek from using distillation techniques. Naveen Rao, the AI vice president at Databricks, noted that learning from competitors is common in the AI sector, much like how car manufacturers study each other’s engines. He stated, "This competition is a real factor, and companies will extract information when possible to gain an advantage."

During a Senate confirmation hearing, Howard Lutnick, nominated by President Trump for Secretary of Commerce, stated that it appears DeepSeek misappropriated US technology and promised to enforce stricter restrictions. He added, "I do not believe that DeepSeek operated above board. I will pursue restrictions rigorously to maintain our lead in AI technology."

Additionally, David Sacks, the White House's AI and cryptocurrency advisor, highlighted concerns regarding DeepSeek's distillation during a recent interview.

DeepSeek has not commented on the allegations against it yet, while OpenAI affirmed its intention to collaborate with the US government to safeguard American technology, although specific methods were not detailed in their response.

The current fears surrounding China’s exploitation of US technology parallel past worries about the semiconductor industry, where the US has already imposed limits on the shipment of specific chips and manufacturing equipment to China and is analyzing the restriction of certain open technologies.

Challenges in Enforcement

Technologists believe that blocking distillation may be more difficult than it appears. One of DeepSeek’s notable breakthroughs is demonstrating that a relatively small number of data samples, less than one million, from a larger, more capable model can significantly enhance a smaller model's performance.

With popular AI tools like ChatGPT having hundreds of millions of users, detecting small data loads can be challenging. Models such as Meta's Llama and offerings from the French startup Mistral can be easily downloaded and applied in private data centers, complicating the detection of violations of their terms of service.

Umesh Padval, managing director at Thomvest Ventures, remarked that stopping model distillation is nearly impossible when open-source models are involved. He said, "Open-source models like Mistral and Llama are accessible to everyone. Additionally, they could potentially locate OpenAI's models through partner customers."

While Meta’s licensing for the Llama model requires users to disclose their distillation practices, DeepSeek’s published paper acknowledged using Llama for certain distilled model versions released recently, but it failed to clarify if they used Meta’s model earlier in the development process. A Meta representative declined to comment on whether DeepSeek violated its terms.

A source familiar with a major AI lab conveyed that stringent identification requirements, similar to those in financial sectors, may be the only way to prevent firms like DeepSeek from distilling US models. However, such measures are not currently established, and while the Biden administration had proposed them, the Trump administration may not support them.

In an effort to restrict Chinese access, Jonathan Ross, CEO of Groq, an AI computing company, stated he had blocked all Chinese IP addresses from his cloud's access, aiming to stop Chinese firms from leveraging the AI models he hosts. Ross acknowledged that such measures may not be fully effective, saying, "People will find ways around it. We have concepts to prevent that, but it will be an ongoing battle. If anyone has a solution, we are eager to implement it."

China, AI, DeepSeek, Technology, US