Aether (Chinese: 以太) is a control model for humanoid robots published by the Chinese robotics company JoyIn, which trades under the registered entity 苏州乐享智能科技有限公司 and sells home robots under the Zeroth brand, covered at Zeroth Robotics. The company named the model on 10 September 2026, two weeks after its demonstration videos began circulating anonymously under the internal codename MVP, which it glossed as Make Veritable People and translated into Chinese as 做个人吧.[1][2][8][11]

The announcement came attached to an open statement from JoyIn founder Guo Renjie (郭人杰), published in Chinese on the company's WeChat account and headed 公开声明：OpenAI，请回答三个问题, meaning "Public statement: OpenAI, please answer three questions". In it Guo accused OpenAI of copying JoyIn's technical framing, its concept of an intelligence conceived as a visitor from beyond the stars, and the design of the Aether website.[1][2][3]

The accusations are JoyIn's alone, and so is every performance figure below. CNBC, which broke the story in English, wrote that it "was unable to independently verify the claims" and noted that several of the concepts Guo points to are existing parts of AI research that companies connect and implement differently. The Next Web wrote that it had not verified the claims either, and separately that it had not verified Aether's performance figures. OpenAI did not immediately respond to CNBC's request for comment.[1][2]

Background

JoyIn was founded in December 2024 by Guo Renjie, previously China executive president of the appliance maker Dreame, and raised a pre-Series A round of close to 500 million yuan led by Ant Group in July 2026. It describes six in-house technology tracks, of which the model layer is one.[12] In an interview published on 3 July 2026, answering a question that framed embodied intelligence as body, cerebellum and brain, Guo said the "brain" was the last of those technologies the company had begun work on, and that JoyIn would release its own model in September.[11] The company's consumer products and funding history are covered on the Zeroth Robotics page.

Claimed specifications

Item	Claim	Attribution
Parameters	4 billion	JoyIn, via the statement as reported by The Next Web and via company disclosure to 硅星人[2][8]
Human video used in training	About 200 hours	JoyIn, same sources[2][8]
Real-robot data used in training	None	JoyIn, same sources[2][8]
Success rate on first attempt	Above 90%, without task-specific training	JoyIn, as reported by The Next Web; CNBC reported the claim as a 90% success rate on first attempt, without the qualifier[1][2]
Zero-shot success rate	80%	The team, as reported by QbitAI on 3 September 2026, a week before the model was named[7]
Training time	Cut by two thirds	Zhu Mingxuan, who led the model's development, to CNBC[1]
Robot platforms	Unitree G1 and AgiBot A3	JoyIn, as reported by The Next Web and 硅星人[2][8]

Zhu Mingxuan told CNBC that she had left the United States humanoid company Figure the previous year, where she had also worked on models for helping humanoid robots mimic human actions, and that JoyIn planned to open-source parts of Aether relating to touch but not the portions related to energy use.[1] 硅星人 reported that JoyIn had recruited model staff from Figure and Skild AI, and that it assembled a model team of about 20 people in March 2026, after which Aether began to take shape.[8] The Next Web also gives the team as 20 people assembled in March.[2]

Architecture as described by JoyIn

JoyIn presents Aether as a bionic neural pathway rather than a policy trained to predict the next action. Its site says the model runs from high-level episodic memory and intent cognition, which it maps to the brain, down to low-level force and position control and reflexes, which it maps to the spinal cord, and that this gives a robot physical intuition and metacognition for the first time. Five capabilities are named: active perception, multimodal understanding, streaming memory, motion dynamics and self-iteration. The same page carries a diagram of the brain regions involved and credits Neurotorium.org as the source of that model.[5]

An architecture page on the same site, at the path /model-0730, labels the stages with brain regions: the superior colliculus for active acquisition of information, the angular gyrus for multimodal perception, the entorhinal cortex for episodic memory encoding, the superior parietal lobule for dynamics constraints and a model of physical consequences, the prefrontal cortex for reasoning and abstraction, the motor cortex for trajectory planning, and the cerebellum for motor execution. Vision, proprioception, spoken commands and touch are shown as inputs, with some pathways marked as activated only when information is explicitly requested.[6]

Writing on 11 September, 硅星人 described three further pieces of the design as JoyIn presents it: an energy formulation that puts task completion, physical constraints, memory and action feasibility into one function so that the robot searches for a direction of action rather than retrieving a memorised one; a self model that sits alongside a world model and tracks what the robot's own body can reach and do, which is the part JoyIn credits for letting one model drive two differently sized machines; and an Evolution Manager that keeps updating knowledge, skills, strategies and what JoyIn calls cognitive structure from real interaction.[8] The know-me page carries the slogan 机器人的ChatGPT时刻，就是抛弃ChatGPT, meaning that the ChatGPT moment for robots is to abandon ChatGPT, and the home page sets Aether against both vision-language-action models and world action models by name.[4][5]

Demonstrations and the anonymous period

The demonstration that started the attention was a ten-minute single-take video that spread through Chinese embodied-AI circles in late August 2026, showing two humanoid robots of different heights cleaning glass, tidying objects, opening a fridge and handling interruptions in what 硅星人 calls a real home environment. JoyIn's statement cites QbitAI's report of it.[2][8]

QbitAI, which had been in contact with the then-unidentified team, reported on 3 September that the team declined to claim the work publicly, telling it that the model is like a person and should speak for itself. The same piece disclosed the codename MVP and carried four further clips supplied by the team, including a robot keeping a cup of water level while being pushed with a stick and steadying falling cans with its other hand, two robots of different heights stretching a bedsheet together, and a robot loading objects onto its own body to avoid a second trip. QbitAI noted that some viewers had asked whether the footage was teleoperated.[7] These are demonstrations selected and released by the vendor, not a benchmark result.

The statement of 10 September 2026

The statement puts three questions to OpenAI covering technical architecture and ideas, the concept of an alien intelligence, and website design.[2][3]

On ideas, Guo says JoyIn presented its framework of an intelligence conceived as a visitor from beyond the stars (星外来客) at a Silicon Valley exchange before OpenAI chief scientist Jakub Pachocki published the essay An Alien Mind on 6 September 2026. The Next Web, reading the letter, gives the date of that presentation as 16 August 2026; CNBC says only that it happened a few weeks earlier. Neither outlet cites an account of the presentation independent of the letter. Guo points to overlaps on recursive self-improvement and on using AI to optimise computing power.[1][2] Pachocki's essay does discuss both: it calls machine recursive self-improvement central to future scientific discovery and says AI will improve the computational substrate itself.[2] 硅星人 adds that JoyIn's site had separated goal alignment from value alignment in July and that Pachocki's essay makes the same separation.[8] CNBC observed that some of the concepts Guo cites are existing parts of AI research more broadly, and linked to an essay on recursive self-improvement published by the Anthropic Institute.[1][2]

On design, Guo compares OpenAI's page for GPT-6 Astra, launched on 3 September 2026, with the Aether site, names particle effects as the core design idea, and describes the result as pixel-level copying in The Next Web's translation.[1][2] The sources disagree about when the Aether site went live. CNBC reports Guo saying it was released two months earlier; Chinese accounts of the letter say more than two months before the Astra page; The Next Web writes that JoyIn says the site went live at the end of July, "more than two months earlier", which is internally inconsistent with a 3 September comparison date.[1][2][8] The Aether site's own robots.txt excludes a path named /model-0730, and that page is the architecture diagram.[6][13] The Internet Archive's first capture of aether.cn in its present form is dated 8 September 2026, which neither confirms nor contradicts a July launch, since the archive holds no captures of the domain between 2021 and that date.[14]

The line most widely quoted from the statement is Guo's charge of distillation, given here in CNBC's translation: "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."[1]

WeChat serves article pages to automated requests behind a verification wall, so the Chinese original was read for this article from screenshots of it that a reader posted to the developer forum linux.do on 10 September 2026, alongside the article URL. Those screenshots show the company's blue-verified account, registered to 苏州乐享智能科技有限公司 under the name 乐享科技 JoyIn, and the closing passage of the statement with its signature.[3][16] Translated from that Chinese text by the author of this article, the closing passage reads in part: "Some things can be imitated, but the whole system that belongs to our underlying design and is built for embodied bodies cannot be copied away. Even against OpenAI, we are entirely confident, because we are the ones who are the open AI." It goes on: "We only hope that competition is transparent and that originality is respected", and "Facts beat concepts, and products beat manifestos." It is signed Guo Renjie, JoyIn, 10 September 2026.[3][16]

Legal action

The same closing passage says, again translated here, that the company will not stay silent in the face of this act of plagiarism, that its legal team has fully started the relevant legal procedures and started litigation, and that JoyIn will use the law to defend the research results belonging to the company and to its team.[3][16] CNBC rendered this as the startup having "started the process of filing a lawsuit". The Next Web wrote that JoyIn's legal team had "fully launched" legal proceedings and started a lawsuit, while also noting CNBC's weaker formulation.[1][2] No court, jurisdiction, defendant filing or case number appears in the statement or in the coverage reviewed for this article, and a search of later reporting up to 13 September 2026 found none.

Reaction and context

OpenAI did not immediately respond to CNBC's request for comment on 11 September 2026.[1] OpenAI's own site blocks direct requests, so its newsroom index was checked through the Internet Archive's capture of 10 September 2026, which carries no item on JoyIn or Aether.[15] That capture is timestamped before CNBC's request for comment on 11 September and so cannot speak for the days after it; in the reporting reviewed for this article, up to 13 September 2026, no response from OpenAI appears.

The statement landed two days after a joint advisory from the United States National Security Agency, Cybersecurity and Infrastructure Security Agency and Federal Bureau of Investigation, numbered AA26-251A and dated 8 September 2026, which accused six China-based AI companies of industrial-scale distillation of United States models: DeepSeek, Moonshot AI, Alibaba, MiniMax, StepFun and Z.ai.[10] JoyIn is not among the companies named in that advisory.[10] CNBC framed the episode as a Chinese startup flipping the distillation claim back at OpenAI.[1][2]

Reporting the same week, CNBC also quoted Garry Tan, chief executive of the accelerator Y Combinator, taking the opposite side of the distillation argument. "I would do nothing" about distillation, Tan told CNBC at Y Combinator's Demo Day, adding: "We could argue that there should be an American distillation regime." Tan also pointed to the copyright status of much of the data used to train frontier models.[9]

Naming

Several unrelated things in AI carry the name Aether. Shanghai AI Laboratory open-sourced a generative world model called AETHER on 23 April 2025, described in the paper "AETHER: Geometric-Aware Unified World Modeling".[17] A separate company called Aether AI, based in San Diego and founded by the University of California, San Diego assistant professor Biwei Huang to work on causal world models, announced a first round of about $20 million led by Matrix Partners China on 17 June 2026.[18] Neither is connected to JoyIn's model.

References

Editing: Aether (JoyIn)