Experts Question Claims That Chinese AI Models Rely on Distillation

The allegations from Kratsios, supported by Treasury Secretary Scott Bessent, suggest that U.S. models are being systematically mined for their proprietary capabilities. This process, known as distillation, involves querying an existing model to generate training data for a new one. However, experts in the field remain unconvinced that distillation alone accounts for the rapid evolution of Chinese AI platforms. Braden Hancock of the Laude Institute notes that the timeline is technically prohibitive, as Fable only became public in July, leaving insufficient time for a comprehensive distillation and training cycle before Kimi K3’s release.

Beyond the logistical hurdles, researchers like Nathan Lambert of the Allen Institute for AI suggest that the industry is shifting toward reinforcement learning, which requires infrastructure and expertise far beyond basic fine-tuning. Distilling frontier-level capabilities via API queries would be prohibitively expensive and slow, failing to yield the performance gains seen in Kimi K3. Critics of the official narrative argue that U.S. policymakers may be underestimating the technical talent within Chinese research teams, who are increasingly capable of independent innovation.

While Anthropic previously accused several Chinese firms, including Moonshot, of deliberate capability extraction based on IP address patterns, the line between distillation and the development of synthetic datasets remains notoriously blurry. Even within the U.S., the practice is not unique; Elon Musk has acknowledged that his company used OpenAI models to develop Grok. The current debate also touches on the black market for restricted Nvidia chips, such as the Grace Blackwell 300, highlighting gaps in global supply chain oversight. Experts like Sam Bresnick argue that without stricter "know your customer" regulations for data centers, tracking the origin of these high-performance models will remain a significant challenge.