US and tech firms debate regulation of AI model distillation amid China concerns
Juste les faits

US and tech firms debate regulation of AI model distillation amid China concerns

Résumé

U.S. officials and leading AI companies are weighing how to address AI distillation, a technique that allows smaller models to learn from larger ones, after Chinese labs released competitive open-weight models.

Google AI leader Jeff Dean described "distillation" as a method for creating smaller, capable models by training them on the outputs of a larger, frontier model. The technique, long discussed among researchers, has recently drawn attention from policymakers and industry leaders because Chinese firms such as Moonshot AI have released open-weight models that appear to use distillation to replicate capabilities of U.S. systems.

White House advisor Michael Kratsios claimed on social media that Moonshot AI distilled Anthropic's Fable model to develop its K3 product, alleging a sophisticated platform designed to evade detection. The accusation reflects broader concerns that distillation could enable intellectual-property theft and give foreign competitors an unfair advantage.

In response, more than 20 technology companies, including Nvidia, Microsoft, Meta and Palantir, issued a joint letter urging lawmakers to avoid "premature restrictions" on open-weight AI models, arguing that distillation is a standard practice for model improvement and validation.

Experts note the policy dilemma. Georgetown researcher Colin Shea-Blymyer said the U.S. government is assessing whether foreign firms are leveraging American model outputs to gain a performance edge. Box CEO Aaron Levie, a signatory of the letter, emphasized that access to the best technology—regardless of origin—is essential for U.S. competitiveness.

Industry voices also highlighted legitimate uses of distillation. Nvidia has employed the technique in training its Llama Nemotron series, and Info-Tech Research Group's Shashi Bellamkonda called it a valuable method for creating cheaper, efficient models. Conversely, Anthropic and OpenAI have prohibited distillation in their terms of service, citing concerns over IP theft and national-security risks.

Anthropic reported that Chinese companies DeepSeek, Moonshot and MiniMax had conducted large-scale distillation of its Claude model using thousands of fake accounts. The firm warned that unchecked distillation could facilitate malicious applications and called for coordinated action among industry, policymakers and the global AI community.

Security firm SecurityPal founder Pukar Hamal said his company would consider using Chinese open-weight models like Kimi K3 after conducting security assessments, noting potential cost savings. Legal commentator Max Pritt observed that while the administration emphasizes protecting corporate IP, it has been less vocal about the broader use of copyrighted content in AI training.

The debate underscores the tension between fostering AI innovation, safeguarding intellectual property and addressing national-security concerns as the technology evolves worldwide.

Source

CNBC

Continuez à lire — gratuitement

Créez un compte gratuit pour suivre l'actualité. Aucune carte requise.