A humanoid robotics company backed by Ant Group has formally accused OpenAI of appropriating its artificial intelligence framework. Guo Renjie, chief executive of JoyIn, released a public letter in Chinese stating that OpenAI’s recently published work constitutes an unaltered distillation of the startup’s proprietary technology.
Guo pointed to the timing of the allegations, noting that JoyIn had presented its “extraterrestrial visitor” model architecture at a Silicon Valley event several weeks prior to OpenAI chief scientist publishing “An Alien Mind” on September 6.
According to a CNBC translation of the statement, Guo expressed skepticism regarding OpenAI’s independence, remarking, “People often say major tech companies have intelligence networks monitoring the whole internet, this time I believe it.” He highlighted specific parallels in core technical strategies, including recursive self-improvement and the use of AI to optimize computational resources.
Additional similarities were drawn between the visual design of OpenAI’s GPT-6 Astra website and the online presence for JoyIn’s Aether model, which Guo asserted was released two months earlier. The company confirmed it has begun legal proceedings to address these concerns. JoyIn states that the Aether model utilizes a perception-based control system rather than text-based inputs, achieving a 90 percent success rate on initial task attempts for humanoid robots.
The allegations arise amid heightened scrutiny of model distillation practices. On Tuesday, a U.S. cybersecurity agency reported that six Chinese entities, including Alibaba and DeepSeek, had distilled models from Anthropic, Google, and OpenAI. Anthropic has previously flagged similar unauthorized replication efforts by Chinese firms.
Zhu Mingxuan, who led the development of the Aether model after leaving U.S. humanoid robot manufacturer Figure last year, told CNBC that the new framework reduced training time by two-thirds. She indicated plans to open-source components related to tactile sensing while keeping energy efficiency modules private.
Industry responses to distillation vary. Garry Tan, chief executive of Y Combinator, stated this month that his accelerator would not intervene in distillation disputes, though he acknowledged that U.S. companies train their models on data protected by copyright law. OpenAI did not immediately respond to a request for comment.
CNBC was unable to independently verify the technical similarities claimed by JoyIn. Some of the contested concepts are recognized elements within broader AI research, though implementation methods often differ between organizations.
JoyIn claims 90% task success with Aether. If their model is so unique, why would OpenAI need to copy it instead of building their own?
Can someone explain what ‘distillation’ actually means here? Is stealing code the same as replicating the architecture? Seems legally blurry.
OpenAI training on copyrighted data is an open secret. Maybe they’re just doing to others what they’ve been accused of for years.
It’s wild that Guo mentioned ‘intelligence networks monitoring the internet.’ Feels like a sci-fi thriller plot rather than a standard IP dispute.
Everyone says this is just distillation, but the visual design similarities are hard to ignore. Did OpenAI copy or just coincidentally converge?