Distillation is now a symmetric move

Distillation is no longer a one-way accusation in frontier AI; it has become a symmetric move, deployed by whoever needs the cover of being wronged. JoyIn, a Suzhou-based humanoid-robotics startup, walked into the opening that a U.S. cybersecurity advisory carved on Sept. 8, when six Chinese firms were named for distilling Western frontier models. Four days earlier, OpenAI's chief scientist published "An Alien Mind," and JoyIn CEO Guo Renjie answered on Sept. 10 with a public letter accusing OpenAI of "direct distillation" of JoyIn's Aether model, the same vocabulary, deployed in reverse.

The mechanism underneath the noise is a vocabulary transfer. When Anthropic, Google, and OpenAI spent two years accusing Chinese labs of distillation, they did not just win a moral argument; they licensed a word. Now a 90%-first-attempt humanoid-control model from a startup whose lead developer, Zhu Mingxuan, left Figure last year, can use that same word against them. CNBC could not independently verify the technical claims, and OpenAI did not respond. The word is doing work the evidence cannot yet do.

Three signals to watch: OpenAI's first on-record response, the specific filings in the lawsuit Guo says JoyIn has begun, and whether the cited overlaps, recursive self-improvement and AI-to-optimize-compute, are concrete enough to litigate. When both sides share a vocabulary, the vocabulary stops describing conduct and starts performing grievance.