Moonshot accused of copying Fable for Kimi K3

Moonshot Accused of Copying Anthropic's Fable

Michael Kratsios, the White House's chief science advisor, recently pointed fingers at Moonshot, a Chinese company, for allegedly copying Anthropic's language model, Fable. According to him, Moonshot purportedly used banned chips for export to China to develop Kimi K3, the largest open-weight language model available. Kratsios described this practice as "large-scale industrial distillation" aimed at stealing American technology and undermining national research. However, he did not provide further details on the sources of his accusations, and Moonshot has not responded to inquiries regarding its training process.

Accusations Resonate in the Industry

Kratsios's tweet resonated with Scott Bessent, the Treasury Secretary, who also expressed concern about the presence of "watermarks" from American models on Chinese models. However, the Treasury Department did not clarify what these watermarks entail. This situation has sparked discussions about the potential ban on Chinese open-weight models, shaking the artificial intelligence sector.

Experts Skeptical About Distillation

Despite these accusations, some experts remain skeptical about the idea that distillation is behind Kimi K3's advanced performance. Braden Hancock, a researcher at the Laude Institute and co-founder of Snorkel AI, expressed doubts about the possibility of creating such a powerful model in such a short time after Fable's public release. According to him, strict distillation would not allow for achieving such a level of performance in just two weeks. Fable has only been publicly available since July 1, making this hypothesis unlikely.

The Role of Reinforcement Learning

Nathan Lambert, an AI researcher at the Allen Institute for AI, also shared his perspective on the matter. In a podcast, he explained that distillation loses its impact as Chinese models approach the technological frontier and training evolves towards reinforcement learning. He argued that if distillation were sufficient, other models like GLM or K3 would have already been caught up by competitors using this method.

Challenges of Distillation and Fine-Tuning

Distillation often involves querying a target model to generate training data. This process can include explicit requests for the model to explain its reasoning. The responses obtained are then used to train a new model through supervised fine-tuning (SFT). However, Lambert emphasizes that SFT becomes less effective as models grow more complex, necessitating advanced techniques like reinforcement learning.

Infrastructure and Cost of Reinforcement Learning

Reinforcement learning requires significant infrastructure, often involving tens of millions of agents to evaluate and adjust model responses. Using a cutting-edge lab's API for this type of training would be extremely costly and could create a time bottleneck. Moreover, there is no guarantee that it would improve the model's performance.

Anthropic's Accusations Against Moonshot

Anthropic has previously accused Moonshot, along with other companies like DeepSeek and MiniMax, of distilling its models. According to Anthropic, millions of exchanges between its models and users identified at these companies have been detected, reflecting a deliberate extraction of capabilities rather than legitimate use. These exchanges were identified through IP addresses and other metadata.

Distillation, a Common Practice

Distillation is not an isolated practice in China. Elon Musk testified that his company SpaceXAI had distilled OpenAI's models to develop Grok, highlighting that this practice is common in the industry. The line between distillation and the development of synthetic datasets is often blurred.

Underestimating Chinese Expertise

Braden Hancock criticized the tendency to underestimate the technical expertise of Chinese teams. He noted that one of the founders of Moonshot was a PhD student at Carnegie Mellon University, emphasizing that these researchers and engineers are legitimate and skilled. According to him, even if American models were to stop, Chinese advancements would continue.

Controversial Access to Nvidia Chips

Kratsios's accusations also include Moonshot's access to advanced Nvidia chips, the Grace Blackwell 300, and servers equipped with GB300 in Thailand. These chips are banned for export to China, but a black market exists. Sam Bresnick, a researcher at the Georgetown Center for Security and Emerging Technology, highlighted the importance of know-your-customer laws for data centers. In May, the founder of Supermicro, an American server manufacturer, was indicted for introducing advanced chips into China.

Towards Stricter Regulation

The Commerce Department under President Joe Biden has proposed federal rules for 2024 regarding know-your-customer regulations in data centers. However, no significant progress was made under Donald Trump. Exporters of advanced chips must ensure that they are used only for approved purposes, but implementing these measures remains a challenge.