White House science adviser Michael Kratsios announced that Moonshot, the Chinese company behind the Kimi K3 large‑language model, built its system by copying Anthropic’s recently released Fable model and by employing Nvidia Grace Blackwell 300 chips that are not cleared for export to China. Kratsios called the alleged “large‑scale, covert industrial distillation” a breach of U.S. research security and said it undermined American innovation.
Kratsios did not disclose the evidence behind his claim, and Moonshot declined to comment on its training methods. Treasury Secretary Scott Bessent added that the administration is detecting “watermarks of our U.S. large language models on many of the Chinese models,” though he offered no details about what those watermarks contain.
Experts doubt rapid distillation explains Kimi K3’s performance
Researchers familiar with AI model development pushed back on the notion that Moonshot could have duplicated Fable’s capabilities within weeks. Braden Hancock, a Laude Institute researcher and co‑founder of Snorkel AI, said the timeline makes sense impossible: “Fable’s only been publicly available since July 1st. You can’t distill that much data, train a model, and release it in two weeks.”
Similarly, Nathan Lambert of the Allen Institute for AI argued that as models approach the frontier, simple supervised fine‑tuning—often called distillation—has diminishing returns. Lambert noted that achieving Fable‑level performance would likely require reinforcement learning from human feedback, a process that demands “tens of millions of agents” and massive compute resources, making it “insanely expensive and potentially a time bottleneck.”
The experts explained that distillation involves systematically querying a target model to generate training data for a new model, sometimes extracting chain‑of‑thought reasoning. While the technique can yield improvements, it does not replace the need for large‑scale reinforcement learning and specialized hardware.
Broader concerns about IP theft and chip smuggling
Anthropic previously accused Moonshot, DeepSeek and MiniMax of systematically extracting its models, citing “millions of exchanges” traced to those companies via IP addresses and metadata. Anthropic did not respond to follow‑up queries about the alleged Fable distillation.
Elon Musk testified earlier this year that his SpaceXAI team distilled OpenAI models to develop its Grok system, underscoring that the practice is not unique to Chinese firms. Hancock emphasized that Moonshot’s founders include a Carnegie Mellon PhD graduate, suggesting a high level of technical expertise.
Kratsios also claimed Moonshot accessed Nvidia Grace Blackwell 300 processors and GB300‑equipped servers located in Thailand—hardware that U.S. export regulations forbid shipping to China. Sam Bresnick, a research fellow at Georgetown’s Center for Security and Emerging Technology, said a black market for such chips exists and referenced a recent indictment of Supermicro’s founder for smuggling advanced chips into China.
President Joe Biden’s Department of Commerce proposed “know‑your‑customer” rules for data centers in 2024, aiming to track who uses high‑end hardware abroad. The proposal stalled under the previous administration, and enforcement remains uneven.
The controversy spotlights a widening rift between the United States and China over AI development, intellectual‑property protection, and the enforcement of export controls on cutting‑edge compute equipment.
Este artículo fue escrito con la asistencia de IA.
News Factory APP - noticias agénticas para impulsar tu SEO y AEO.