Michael Kratsios, a science advisor to the White House, has raised concerns regarding Moonshot, a Chinese company that developed the Kimi K3, the largest open-weight language model currently available. He alleges that the company replicated Anthropic’s Fable LLM and utilized chips not authorized for export to China in the process.
Kratsios emphasized the seriousness of industrial espionage aimed at appropriating U.S. technology and undermining American research efforts. Amidst ongoing discussions about potentially restricting Chinese open-weight models, he pointed out that Moonshot had not clarified its training methods, and he refrained from providing additional information on his claims.
His remarks echoed those made by Treasury Secretary Scott Bessent, who indicated that markers of U.S. language models have been found in various Chinese models, though the specific nature of these markers remains ambiguous with no response from the Treasury Department on this matter.
Experts, however, express doubts about the idea that distillation—a method involving querying an LLM to understand and replicate its functions—was solely responsible for the impressive performance of Kimi K3.
“I don’t believe that a model of this quality can be produced so quickly using only distillation from Fable,” remarked Braden Hancock, a researcher at the Laude Institute and co-founder of Snorkel AI. He noted that Fable became publicly accessible only recently, and realistically, the timeline does not support the claim that such extensive data can be distilled and a model trained within two weeks.
Nathan Lambert, an AI researcher at the Allen Institute for AI, mentioned that the significance of distillation is diminishing as Chinese models evolve closer to state-of-the-art technologies, implying that superior training methods like reinforcement learning are becoming necessary. He argued that if distillation were sufficient, many would easily replicate advanced models, which has not occurred through conventional supervised fine-tuning.
The distillation process relies on detailed interactions with the target model to generate data for further training. This may involve the model explaining its reasoning to better understand problem-solving approaches. The resulting exchanges can then be used for supervised fine-tuning, a method through which a new model can claim to be mimicking a well-known model.
Lambert suggested that the impact of supervised fine-tuning is waning as models become increasingly intricate. To replicate the capabilities of a model like Fable, methods involving reinforcement learning would likely be needed. This involves an advanced approach where a larger model assesses the responses of a smaller model and adjusts its training based on those evaluations.
Such sophisticated techniques also require substantial infrastructure, as large runs of reinforcement learning may necessitate millions of agents. Using an existing advanced model’s API for this could be prohibitively expensive and time-constraining, especially given the slow processing speeds of such models, which may not yield significant performance enhancements.
It is suspected that prior leading models may have influenced Kimi; Anthropic previously accused companies including Moonshot of systematically extracting capabilities from its models, discovering anomalous user interactions linked to these organizations through data analysis. Such queries were characterized as suspiciously focused on gaining model capabilities rather than typical usage.
The practice of distillation is recognized as common across the AI industry, not exclusive to China. Elon Musk disclosed earlier this year that his company distanced OpenAI models for their own development, indicating that this process is prevalent in AI development.
Hancock emphasized the technical abilities of Chinese teams, noting that one of Moonshot’s founders has a PhD from a prestigious university. He suggested that even if American contributions stagnate, China’s advancements in AI would persist, disputing the notion that they solely rely on external technologies.
The allegation that Moonshot acquired advanced Nvidia chips, which are prohibited for export to China, complicates the matter further. Some sources indicate a black market for such components exists. In light of this, there have been calls for stricter regulations to monitor companies using cutting-edge training facilities responsibly.
In response to these concerns, the Department of Commerce has proposed regulatory measures for data centers but progress has been stalled. Exporters of advanced chips must ensure compliance with export regulations, ensuring they are utilized for sanctioned purposes.
