A Chinese humanoid robotics startup is accusing OpenAI of directly copying its AI model architecture, joining a growing chorus of complaints about what companies call "distillation" — the practice of harvesting outputs from advanced systems to train cheaper replacements.
Guo Renjie, CEO of Suzhou-based JoyIn, published a public letter Thursday alleging that OpenAI's GPT-6 Astra, launched Sept. 3, bears unmistakable technical and design similarities to JoyIn's Aether model, which the company says it had been developing for months. Guo said the resemblance went beyond vague conceptual overlap.
"People often say major tech companies have intelligence networks monitoring the whole internet — this time I believe it," Guo said in a statement translated by CNBC. He said JoyIn had publicly presented its Aether framework in Silicon Valley several weeks before OpenAI's Chief Scientist published a paper titled "An Alien Mind" on Sept. 6.
Guo pointed to shared architectural approaches, including recursive self-improvement and the use of AI to optimize computing power. He also highlighted parallels in branding and design: both OpenAI's GPT-6 Astra landing page and JoyIn's Aether website feature outer-space imagery and naming, JoyIn claims it first released its site two months ago.
CNBC was unable to independently verify the allegations. Several of the concepts cited are established components of AI research broadly, though implementation details vary between companies. OpenAI did not immediately respond to a request for comment. Guo confirmed Thursday that JoyIn has begun filing a lawsuit.
The dispute arrives against a backdrop of heightened U.S. scrutiny of Chinese AI firms. On Tuesday, a U.S. cybersecurity agency said six Chinese companies — including DeepSeek and Alibaba — had distilled models from Anthropic, Google and OpenAI. Anthropic has repeatedly raised alarms about unauthorized model extraction by Chinese firms seeking to boost their own capabilities.
Garry Tan, chief executive of Y Combinator, told CNBC this month he would take no action on distillation, joining other U.S. executives in noting that American companies had trained their models on data covered under copyright law — a legal argument that could complicate enforcement.
JoyIn's Aether model, announced Thursday, claims a perceptive rather than text-based approach to robotic control that enables humanoids to complete tasks with a 90% success rate on first attempt. Zhu Mingxuan, who led the model's development, left U.S. humanoid robotics company Figure last year after working on models designed to help robots mimic human movement. She told CNBC earlier this week that Aether reduced training time by two-thirds. JoyIn plans to open-source portions of the model related to touch feedback but will retain proprietary controls over energy-optimization components.
Achieving reliable, human-level performance across diverse tasks remains a significant challenge for the humanoid robotics sector, even as the United States and China compete aggressively to establish dominance.
