The chief executive of an Ant Group-backed humanoid robotics startup has accused OpenAI of directly copying elements of its artificial intelligence research and product design, escalating a growing debate over AI “distillation” as U.S. and Chinese companies race to develop more capable models for robots.
Guo Renjie, CEO of Suzhou-based JoyIn, published an open letter to OpenAI in Chinese on Thursday questioning what he described as striking similarities between the startup’s recently presented technology and OpenAI’s latest AI research.
“People often say major tech companies have intelligence networks monitoring the whole internet, this time I believe it, this is a direct distillation of us without any modifications,” Guo said in the statement, according to a CNBC translation.
Register for the next Tekedia Mini-MBA.
Register for Tekedia AI in Business Masterclass.
Join Tekedia Capital Syndicate and co-invest in great global startups.
Some of the concepts cited by Guo, including recursive self-improvement and the use of AI to optimize computing resources, are already part of broader AI research. Companies can also arrive at similar approaches independently while implementing them in substantially different ways.
Guo said JoyIn had publicly presented its “extraterrestrial visitor” AI model framework in Silicon Valley several weeks before OpenAI Chief Scientist published an essay titled “An Alien Mind” on Sept. 6.
He argued that the two approaches shared similarities in their underlying technical concepts, particularly the use of recursive self-improvement and AI systems designed to optimize computing power.
Guo also pointed to similarities in the visual presentation of the companies’ products. He said OpenAI’s GPT-6 Astra webpage uses an outer-space-inspired design similar to JoyIn’s Aether model website, which he said had been released two months earlier.
JoyIn has begun the process of filing a lawsuit, Guo said, potentially turning what began as a public accusation into a legal dispute over intellectual property and the boundaries of independent AI development.
Distillation Becomes a New Fault Line in AI Race
The allegations arrive as model distillation has become one of the most contentious issues in the global AI industry.
Distillation generally involves using the outputs or behavior of a more capable model to help train or improve another system. It can allow developers to reproduce aspects of a frontier model’s capabilities at lower cost, although determining whether a particular system was improperly derived from another model can be technically and legally difficult.
Anthropic has repeatedly accused Chinese companies of using unauthorized distillation of its models to improve their own AI systems.
On Tuesday, a U.S. cybersecurity agency said six Chinese companies, including DeepSeek and Alibaba, had distilled models from systems developed by Anthropic, Google and OpenAI.
The JoyIn allegation introduces another dimension to that dispute because it involves a Chinese robotics company claiming that an American AI developer borrowed concepts in the opposite direction. That does not establish that OpenAI engaged in improper distillation. The underlying concepts cited by Guo are sufficiently broad that similarities alone would not demonstrate that one company copied another’s proprietary technology.
The dispute nevertheless illustrates how difficult it is becoming to draw clear boundaries around originality in AI. Researchers and companies increasingly build on similar ideas, publish technical concepts publicly and train systems using vast quantities of information available across the internet.
As AI development accelerates, the question is shifting from whether companies are influenced by one another to whether they have crossed a line by reproducing proprietary model behavior, architecture, training methods, or product designs without authorization.
JoyIn Takes Its AI Fight Into Humanoid Robotics
JoyIn’s Aether model is designed for humanoid robots and uses what the company describes as a perceptive, rather than primarily text-based, approach to robotic control.
Guo said the system allows humanoid robots to complete tasks with a 90% success rate on their first attempt.
Zhu Mingxuan, who led Aether’s development, said she left U.S. humanoid robotics company Figure last year, where she worked on models designed to help humanoid robots reproduce human actions.
Zhu told CNBC earlier this week that Aether had reduced training time by two-thirds. She also said JoyIn plans to open-source parts of the model, including technology related to touch, while keeping portions concerning energy use proprietary.
The decision to open-source some components could help JoyIn attract researchers and developers to its platform while retaining control over areas it considers commercially sensitive. It also reflects the broader strategy emerging across the robotics industry.
Humanoid companies increasingly need advances in both physical hardware and AI models capable of interpreting environments, planning actions, and responding to changes in real time. That makes model development a critical competitive advantage.
Humanoid robotics remains an especially difficult test for AI because a system must translate intelligence into physical action. A model that performs well in text or image benchmarks can still struggle with balance, dexterity, perception, touch, energy management, and unpredictable real-world environments.
The ability to reliably complete physical tasks therefore depends on more than simply increasing model size or training data.
The competition between U.S. and Chinese developers is intensifying as both countries attempt to establish leadership in AI and physical robotics. Companies are seeking ways to reduce training costs, improve robot performance and shorten development cycles, while increasingly treating model architecture and training techniques as valuable intellectual property.
That environment is likely to produce more disputes over what constitutes inspiration, legitimate research, reverse engineering, or unauthorized copying.
The JoyIn dispute also highlights an uncomfortable feature of the AI industry’s development. Companies frequently publish research, demonstrate products publicly and release portions of their technology as open source because visibility can attract talent, customers and developers. The same openness can make it difficult to prevent competitors from learning from publicly disclosed ideas.
Garry Tan, chief executive of startup accelerator Y Combinator, told CNBC this month that he would “do nothing” about distillation, while noting that U.S. AI companies themselves have trained models on data covered by copyright law.
That argument points to the broader inconsistency surrounding the debate. U.S. companies have raised concerns about competitors learning from their models, while facing their own legal and ethical questions over the data and intellectual property used to train those systems.
For now, JoyIn’s claims against OpenAI remain allegations rather than established findings. But the dispute could grow wider if the startup provides technical evidence showing that proprietary elements of Aether were reproduced rather than merely similar concepts emerging independently.
The issue is likely to become more important as AI moves deeper into robotics. When models control physical machines, the value of proprietary training methods and real-world interaction data can become substantially greater, potentially making intellectual-property disputes more consequential.