Chinese Robotics Startup JoyIn Claims OpenAI Copied Its Aether Model and Website Design

JoyIn Accuses OpenAI of Copying Technical Concepts

Chinese humanoid robotics startup JoyIn has accused OpenAI of copying its ideas and website design, escalating tensions in the U.S.-China tech race. CEO Guo Renjie published an open letter in Chinese on Thursday, claiming OpenAI directly copied the company's work without modifications. "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 stated1

1

.

The Suzhou-based startup, backed by Ant Group, presented its "extraterrestrial visitor" AI model framework in Silicon Valley weeks before OpenAI Chief Scientist Jakub Pachocki published "An Alien Mind" on September 6. Guo questioned similarities in core technological approaches, specifically citing recursive self-improvement and AI-optimized computing power concepts2

2

. OpenAI did not immediately respond to requests for comment, and CNBC was unable to independently verify the claims.

Website Design Similarities Spark Legal Action

Beyond technical concepts, Guo Renjie highlighted what he described as "pixel-level" copying between OpenAI's GPT-6 Astra webpage and JoyIn's Aether model website. The company claims its site went live at the end of July, more than two months before OpenAI launched Astra on September 32

2

. The letter specifically names particle effects as the core design idea allegedly copied.

JoyIn has started the process of filing a lawsuit, with the company stating its legal team has "fully launched" legal proceedings1

1

. This legal action represents a reversal of typical intellectual property disputes in AI, where U.S. agencies have repeatedly accused Chinese companies of unauthorized model distillation from American AI systems.

Aether Model Claims 90% Success Rate for Humanoid Robots

In response to the controversy, JoyIn brought forward the public announcement of its Aether model, a humanoid robot control model that claims breakthrough performance. The company summarizes Aether with three key numbers: 200 hours of human video training data, 0 data from real robots, and 4 billion parameters2

2

.

JoyIn claims the Aether model achieves a 90% success rate on first attempt across various tasks without task-specific training, using what it describes as a perceptive rather than text-based approach to robotic control1

1

. Zhu Mingxuan, who led the model's development after leaving U.S. humanoid company Figure last year, told CNBC that Aether reduced training time by two-thirds. The model runs on robots from Unitree and AgiBot, with a ten-minute demonstration reportedly filmed in a single take.

Technical Architecture and Open-Source Plans

The Aether model features an architecture modeled on brain regions, from the prefrontal cortex to the spinal cord. JoyIn lists five core capabilities: active perception, multimodal understanding, streaming memory, motor dynamics, and self-iteration2

2

. The 20-person team assembled in March developed this humanoid robot control model to address the persistent challenge of getting humanoids and AI models to perform as intelligently as humans across varied tasks.

Zhu indicated plans to open-source portions of the model related to touch and active perception, though components related to energy use will remain proprietary1

1

. This selective approach to open-sourcing reflects broader strategic considerations in AI innovation, where companies balance collaboration with competitive advantage.

Model Distillation Debate Intensifies

The accusations arrive amid heightened scrutiny of model distillation practices. On September 8, U.S. cybersecurity agencies including the NSA, CISA, and FBI accused six Chinese companies—including DeepSeek and Alibaba—of distilling models from Anthropic, Google, and OpenAI1

1

. U.S.-based Anthropic has repeatedly flagged unauthorized distillation of its models by Chinese companies seeking to improve their own AI capabilities.

However, some industry voices question the moral high ground. Garry Tan, CEO of Y Combinator, told CNBC he would "do nothing" about distillation, joining others who point out that U.S. companies have trained their AI models on copyrighted data1

1

. This perspective highlights the complex intellectual property landscape surrounding AI development.

Existing Concepts Complicate Claims

While Guo Renjie's letter emphasizes similarities, CNBC noted that some cited concepts are existing parts of AI research more broadly. Recursive self-improvement, for instance, appears in work from the Anthropic Institute published in June, months before either company's recent announcements2

2

. Pachocki's essay traces OpenAI's work on reasoning models to mid-2023, when a research project called "RLSlow" produced initial results.

The challenge lies in how companies connect and implement these concepts differently. Pachocki wrote that "machine recursive self-improvement will be at the very core of future scientific discovery" and that OpenAI focuses on this because it's "the only way to remain at the frontier"2

2

. Whether JoyIn's specific implementation constitutes protected intellectual property remains to be determined through legal proceedings, as the U.S.-China tech race continues to generate friction over AI innovation and ownership.