JoyIn’s OpenAI “Distillation” Claim: What the Startup Alleged

JoyIn, a Chinese humanoid-robotics startup, has accused OpenAI of copying both its technical direction and visual presentation. CEO Guo Renjie characterized the alleged overlap as a “direct distillation” of JoyIn’s work. The accusation has not been independently verified, OpenAI had not responded to CNBC’s request for comment at publication, and no court has ruled on the matter. 2

What JoyIn says OpenAI copied

In a Chinese-language open letter dated September 10, Guo said JoyIn had publicly demonstrated an “extraterrestrial visitor” AI-model framework in Silicon Valley several weeks earlier. He argued that OpenAI chief scientist Jakub Pachocki’s September 6 essay, An Alien Mind, later reflected similar directions—particularly recursive self-improvement and using AI to optimize computing. 216

Guo also pointed to a design comparison. He said OpenAI’s GPT-6 Astra webpage, which reports described as launching on September 3, resembled the outer-space styling of JoyIn’s Aether site, which he said had been published two months earlier. 2

JoyIn said its legal team had begun the process of filing a lawsuit. That is different from a lawsuit having been filed, heard, or decided. 2

Why the allegation is unproven

The available reporting does not establish that OpenAI accessed JoyIn’s materials, copied protected material, or violated any legal obligation. CNBC said it could not independently verify either the technical or design claims. 2

The ideas highlighted by JoyIn are also not, on their face, proof of copying. Recursive self-improvement and compute optimization are broadly discussed AI-research directions. Similar framing, a shared space motif, or overlapping ambitions do not by themselves demonstrate infringement, trade-secret misuse, or improper model distillation.

OpenAI’s An Alien Mind does discuss model oversight and notes that preventing distillation was a secondary reason for one of its design choices, but that reference alone does not corroborate JoyIn’s claim. 16

What JoyIn’s Aether model is supposed to do

JoyIn announced Aether on September 10 as a control model for humanoid robots. The company describes it as perception-driven rather than primarily text-driven, with the aim of helping robots learn and carry out physical tasks more effectively. 2

JoyIn’s reported product claims include:

• 90% first-attempt task completion
• A two-thirds reduction in training time
• Plans to open-source tactile-sensing-related components
• Retention of energy-use-related components as proprietary technology 2

These performance figures and product characterizations come from JoyIn, not from an independent benchmark or audit. They should be read as company claims rather than verified measures of Aether’s real-world performance.

A reversal in the U.S.–China distillation debate

JoyIn’s allegation is notable because it reverses a prominent line of criticism in the U.S.–China AI competition. In a September advisory, the FBI, NSA, and CISA alleged that six China-based companies—DeepSeek, Moonshot AI, Alibaba, MiniMax, StepFun, and Z.AI—had systematically extracted proprietary capabilities from U.S. frontier models through industrial-scale distillation campaigns. The advisory named variants of Claude, GPT, Gemini, and Grok among the targeted systems.

The U.S. agencies distinguished those allegations from legitimate uses of distillation in AI research, describing the alleged activity as targeted extraction of restricted proprietary capabilities.

China rejected the accusations, arguing that the claims lack factual and legal basis and that model distillation is a widespread practice. China’s Commerce Ministry also warned of countermeasures if the allegations are used to suppress Chinese AI companies.

There is no universal policy consensus on the issue. Y Combinator CEO Garry Tan said he would “do nothing” about model distillation and suggested the U.S. could consider an American distillation regime, a view that contrasts with calls for tighter protections by frontier-model companies and U.S. security agencies.

What “distillation” means—and why facts matter

In its ordinary technical sense, knowledge distillation involves training one model to imitate outputs or capabilities associated with another, often more capable model. It can be a standard technique for building smaller or cheaper systems. The controversy arises when a company is alleged to use access to a proprietary model at scale to reproduce restricted functionality or bypass contractual and technical safeguards.

That distinction is central to both the U.S. government’s allegations and JoyIn’s counter-accusation. Whether conduct is improper depends on evidence of access, scale, copying, circumvention, applicable terms, and the law in the relevant jurisdiction. A resemblance between concepts or webpages is not enough on its own.

For now, JoyIn’s complaint is best understood as an unverified public allegation—and as a sharp illustration of how contested the word “distillation” has become in the global AI race.