Knowledge distillation is a training technique for compressing a larger model's capabilities into a smaller one. As a competitive accusation in the AI industry, it is the claim that one lab trained on another's outputs and passed the result as original work. JoyIn, an Ant-backed Chinese humanoid startup, is now directing that question toward OpenAI: the company's CEO published an article in Chinese alongside the release of JoyIn's new robotics model, raising questions about similarities between JoyIn's work and OpenAI's published research.
The direction of the claim
The framing the CEO chose is a reversal. Rather than fielding a distillation accusation, JoyIn's CEO is pointing to similarities between JoyIn's new robotics model and what OpenAI has published, placing the question of influence on OpenAI's side.
The article was written in Chinese. Which specific sections of OpenAI's research the CEO identified as comparable to JoyIn's model, and the technical character of the overlap, have not been detailed publicly.
JoyIn is backed by Ant Group and operates in the humanoid robotics space. The CEO published the comparison at the same time as a product launch, placing JoyIn's new robotics model at the center of a dispute over technical lineage at the moment of its release.