On January 29, Top White House advisers expressed concern over potential unauthorized gains made by DeepSeek through the use of a technique referred to as "distillation." This method involves one AI system learning from another, which Silicon Valley insiders suggest could be challenging to prevent.
DeepSeek recently introduced a new AI model that appeared to rival the capabilities of competitors, potentially utilizing knowledge from U.S. models without incurring associated costs. This approach, known as distillation, allows newer models to benefit from the investments made in developing older models by transferring their learnings.
Although this distillation technique is common in the AI field, it may violate the terms of service of certain U.S. tech companies. DeepSeek has been scrutinized for potentially using this method improperly to replicate U.S. AI models.
Naveen Rao of Databricks noted that learning from competitors is common practice in the AI industry, akin to how automakers study each other's technology. Despite this, concerns have been raised by officials such as Secretary of Commerce nominee regarding DeepSeek's alleged misappropriation of U.S. AI technology, with a commitment to impose restrictions.
The White House's AI and crypto official also expressed worries about DeepSeek's distillation practices. OpenAI announced its intention to collaborate with the government to safeguard U.S. technology without providing specific details on their plans.
The ongoing apprehension in Washington mirrors past actions restricting the export of certain technologies to China, mirroring the current concerns with DeepSeek's approach to AI model development.
Experts believe that blocking distillation practices, particularly with open-source models available, may be challenging due to difficulties in detection and enforcement. While some companies like Meta have strict terms of service regarding distillation, ensuring compliance remains complex.
Suggestions to curtail distillation include implementing stringent identification requirements for those accessing AI models, similar to financial industry standards. However, definitive measures are yet to be established, with differing opinions on regulatory approaches under the Biden and Trump administrations.
Despite efforts to restrict access to AI models like blocking Chinese IP addresses, industry leaders acknowledge the need for innovative solutions to effectively prevent unauthorized replication. The continuous evolution of technology may require a dynamic approach to safeguard intellectual property in the AI sector.