White House claims Moonshot copied Anthropic's Fable—but experts doubt it

White House science advisor Michael Kratsios published a statement on social media alleging that Moonshot, the Chinese company behind the open-weight LLM Kimi K3, achieved its capabilities by copying Anthropic's Fable LLM while using advanced chips that are legally barred from export to China. Kratsios described the alleged activity as "large-scale, covert industrial distillation aimed at stealing proprietary U.S. technology and undermining American research." He did not provide detailed evidence or respond to follow-up queries. His remarks echoed earlier comments by Treasury Secretary Scott Bessent, who stated that "we are finding watermarks of our U.S. large language models on many of the Chinese models," though Bessent did not clarify what those watermarks represent, and the Treasury Department did not respond to inquiries.

However, leading AI researchers have cast doubt on the distillation narrative. Braden Hancock, a researcher at the Laude Institute and co-founder of Snorkel AI, told TechCrunch: "I don't think you get a model this strong and this quickly on the heels of Fable doing strictly distillation. There's just not even frankly time, right? 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." Nathan Lambert, an AI researcher at the Allen Institute for AI, offered a complementary technical argument: distillation via supervised fine-tuning becomes less effective as models mature and developers shift to reinforcement learning. He explained that frontier-level distillation would require having a large model grade a smaller model's responses and adjust accordingly—a process demanding tens of millions of agents. Running such a system against a frontier model's API "would be insanely expensive and potentially... a time bottleneck." The historical context supports this skepticism: Anthropic disclosed earlier in 2024 that it had detected millions of systematic queries against its models from IP addresses and metadata linked to Moonshot, DeepSeek, and MiniMax, reflecting what Anthropic characterized as "deliberate capability extraction." Yet despite that prior distillation activity, the company did not respond to queries about whether the same pattern occurred with Fable specifically.

The chip access allegation constitutes a separate and more concrete claim. According to Kratsios' remarks, Moonshot obtained Nvidia Grace Blackwell 300s and accessed GB300-equipped servers in Thailand. Both of these are subject to U.S. export controls and legally prohibited from sale to China. Sam Bresnick, a research fellow at Georgetown's Center for Security and Emerging Technology, noted that a black market for advanced chips exists; in May, the founder of Supermicro, a U.S. server builder, was indicted for smuggling advanced chips into China. Bresnick advocated for stricter enforcement: "I am a proponent of know your customer laws for data centers across the world. If you are letting a company conduct huge training runs on your state-of-the-art hardware, there needs to be a reporting mechanism for who that company is and what they're doing." The Biden administration proposed federal know-your-customer rules for data centers in 2024, but no further progress has been reported under the Trump administration.

Researchers also contextualized Moonshot's capabilities within the broader landscape of Chinese AI development. Braden Hancock countered the implication that Moonshot was merely copying, saying: "[I]n general, Americans are understating the technical expertise of these Chinese teams. One of the founders of Moonshot was a CMU PhD student. These are legitimate researchers and engineers doing solid work." He added that even if American models "ground to a halt, I think China's progress would slow, but would still continue. They're not just riding coattails here." The distinction between distillation and legitimate synthetic dataset generation also complicates the narrative; Elon Musk testified earlier this year that SpaceX's AI division distilled OpenAI models to develop Grok, and the industry views distillation as a widespread practice, not unique to Chinese firms.

The allegation from White House science advisor Michael Kratsios represents a high-level assertion of technology theft, but the technical community's skepticism reveals a gap between political framing and engineering reality. Distillation—systematically querying a model to extract its capabilities—is a known technique, and Anthropic itself disclosed in 2024 that Moonshot, DeepSeek, and MiniMax had engaged in systematic queries through its APIs, detected via IP address and usage pattern anomalies. However, researchers emphasize that distillation's effectiveness has diminished as models have advanced. Braden Hancock and Nathan Lambert both argue that the 10-week window between Fable's public launch (July 1st) and Kimi K3's release is too narrow for traditional distillation-based training and deployment, and that modern frontier model capabilities depend on reinforcement learning techniques—which require tens of millions of agents and would be prohibitively expensive to run against a commercial API.

The chip access allegation adds a separate but entangled dimension. Moonshot is said to have obtained Grace Blackwell 300s and accessed GB300-equipped servers in Thailand, both subject to U.S. export controls. This raises questions about training infrastructure rather than training methodology; advanced chips accelerate development but do not by themselves explain how a team could distill and train a competitive model in two weeks. The broader policy question is whether greater transparency at data centers—through know-your-customer rules proposed (but not advanced) by the Biden administration—could prevent future circumvention. The Trump administration has not continued that regulatory effort, leaving a potential enforcement gap.

No comments yet. Be the first to share your thoughts!