Experts say exploiting Anthropic’s Fable isn’t how Kimi K3 got so good

Michael Kratsios, a science advisor at the White House, has raised eyebrows regarding the development of Kimi K3, an advanced open-weight language model created by the Chinese firm Moonshot. Kratsios claims that the model was built by copying Anthropic’s Fable language model, utilizing chips that are not permitted for export to China. He stated, "Large-scale, covert industrial distillation aimed at stealing proprietary U.S. technology and undermining American research is unacceptable," amidst ongoing discussions about potentially banning Chinese open-weight models affecting the AI landscape. Despite the serious allegations, Moonshot has yet to respond to inquiries about its training methods, and Kratsios did not disclose further details backing his claims. His statements reflect concerns echoed by Treasury Secretary Scott Bessent, who mentioned that U.S. large language models' characteristics are being found in many Chinese counterparts, a situation deemed unacceptable. However, the nature of these "watermarks" remains unclear, and the Treasury Department did not provide clarification when contacted. Experts in the field are skeptical that distillation—the process of querying a model to replicate its capabilities—could solely account for the remarkable performance of Kimi K3. Braden Hancock, a researcher at the Laude Institute and co-founder of Snorkel AI, pointed out, "A model this powerful and developed so quickly cannot be attributed to mere distillation, especially since Fable has only been publicly available for a brief period." Nathan Lambert, an AI researcher at the Allen Institute for AI, supported this skepticism in a recent podcast, suggesting that as Chinese models advance, the effectiveness of distillation diminishes. He stated, "If distillation were the key, we would see others quickly catching up to models like GLM or K3, but that has not occurred with just supervised fine-tuning." The distillation process typically involves systematically querying a target model to create data for post-training. This may include asking the model to explain its reasoning or using its responses for supervised fine-tuning training. Lambert noted that while fine-tuning helps a model adopt better behaviors, the advantages of this method are waning as models grow more complex. To replicate capabilities similar to Fable’s, reinforcement learning techniques would likely be necessary, requiring significant infrastructure and resources. Anthropic has previously accused Moonshot, among others, of systematically distilling its models, claiming to have identified millions of interactions between its models and users at those companies, suggesting intentional capability extraction rather than normal usage. Distillation practices are not exclusive to Chinese companies; they are common across the AI industry. For instance, Elon Musk testified that SpaceXAI used OpenAI models to develop Grok, indicating that this practice is widely accepted in the field. Hancock noted the technical expertise of Chinese teams, emphasizing that they have legitimate researchers and engineers capable of significant advancements, independent of American models. The situation is further complicated by Kratsios’ comments regarding Moonshot's access to advanced Nvidia chips, which are restricted from export to China. The existence of a black market for these chips has been noted, raising concerns about compliance with export regulations. Experts like Sam Bresnick, a research fellow at Georgetown, advocate for stricter reporting mechanisms for companies using advanced training hardware, highlighting the importance of transparency in AI development. As the debate continues, it remains to be seen how these allegations will impact the future of AI development and international relations in the tech sector.