EastVoice · AI · 2026-10-26 · Issue #109

Editor's Note: A robot skills contest dropped humanoids into a replica high-voltage substation and let them work to State Grid's real standards. Qiyuan Robotics rolled out its 5,000th personal robot, XAG told the FT it wants a robot on every farm, and AgiBot, Xpeng and Tesla all penciled their humanoid deliveries into the same year. On models, Kimi K3.1 Pro landed in a global top four with the top agent score.

Commentary: China has the factories, the pilots and the deadlines. What it still cannot show is a single site where a robot pays for itself without a subsidy or a camera crew.

---

I. Humanoid Robots

1. A robot contest in a substation splits embodied AI into two schools

• Summary: Humanoids were pushed into a 1:1 replica of an ultra-high-voltage substation and given 30 minutes to read meters, run infrared temperature checks, flag equipment faults and squeeze through a 50 cm aisle, all to State Grid's live operating standards. Huxiu reads the event as the clearest split yet: US teams generate physical experience in code, Chinese teams feed on factory-floor data.
• Source: Huxiu · 2026-10-25
• Editor's Take: Simulation fidelity is the wrong thing to argue about. Whoever owns the messy site data writes the acceptance test, and the acceptance test is the contract.

2. Qiyuan Robotics rolls out its 5,000th personal robot

• Summary: Qiyuan Robotics (启元机器人) said its 5,000th personal robot came off the line, with daily capacity now above 200 units.
• Source: ITHome · 2026-10-25
• Editor's Take: Five thousand consumer units is a manufacturing milestone, not a deployment one. The number that matters is how many are still switched on in a month.

3. Every humanoid maker has picked 2027, and none of them coordinated

• Summary: In ten days, a dozen competing robot companies landed on the same delivery year. Musk put external Optimus shipments at late 2027, He Xiaopeng put IRON's commercial launch at 2027, and AgiBot's Deng Taihua named the same target for volume production.
• Source: TMTPost · 2026-10-25
• Editor's Take: When the whole field picks one year, it stops reading as a roadmap and starts reading as a fundraising calendar.

4. XAG wants a robot on every farm

• Summary: XAG (极飞科技), the SoftBank-backed agricultural drone maker, told the FT it wants every farmer to own a robot, pitching unmanned crop-spraying against a shrinking farm workforce and rising temperatures. By revenue, XAG sits second only to DJI.
• Source: ITHome · 2026-10-25
• Editor's Take: Agriculture is the rare autonomy market that pays without a regulatory waiver. XAG's real fight is with DJI's distribution network, not with anyone's airframe.

5. Beijing raises the bar for robot listings

• Summary: After regulators floated tougher IPO standards for robot companies, the filings kept coming. Songyan Dynamics (松延动力) was reported to have filed confidentially for a Hong Kong listing targeting several hundred million dollars, possibly by 2027; AgiBot (智元) was earlier linked to a ¥1-1.5 billion round.
• Source: Huxiu · 2026-10-25
• Editor's Take: Tightening the listing rules years before the revenue arrives prices the paper valuations, not the robots. The order books have not slowed.

6. Robot founders are burning incense while the IPO window narrows

• Summary: An employee at an embodied-AI company filing for Hong Kong's 18C route said they were still waiting to cash out and watching the valuation shrink with every delay. Huxiu describes engineers on six-figure salaries tracking capital markets between test runs.
• Source: Huxiu · 2026-10-25
• Editor's Take: Options are the retention mechanism of this entire sector. When the exit stops looking real, the salary alone will not hold a robotics team.

II. Autonomous Driving

1. Tesla is building its own AI map, and it ships with FSD v15

• Summary: On Tesla's earnings call, vice president Ashok Elluswamy said the company is building an AI-based navigation system that will arrive with FSD v15, blending fleet-collected data with per-driver route planning and voice control.
• Source: ITHome · 2026-10-25
• Editor's Take: Cutting the map vendor is the actual announcement. Owning the routing layer turns every car into a data source Tesla no longer has to license.

2. Musk pushes a SpaceX-Tesla merger

• Summary: The FT reported that Musk is backing a possible merger of SpaceX and Tesla, folding most of his businesses into one listed company. Any deal still needs an independent board committee, shareholder sign-off and regulatory approval.
• Source: Huxiu · 2026-10-25
• Editor's Take: A rocket company inside a car company is a capital-markets move wearing a technology costume. The governance questions are the whole story.

3. Xpeng's "laodeng" gaffe outlives the apology

• Summary: At the MONA L05 pre-sale launch on Oct 15, Xpeng put young product managers on stage and leaned hard on internet slang. Repeated use of "laodeng," a put-down aimed at older people, blew up; the staffer apologized, executives defended him in long posts, and the company offered a ¥1 million bounty against claims of manufactured outrage.
• Source: TMTPost · 2026-10-25
• Editor's Take: Ten days of damage control over one word is what happens when a brand tries to sound younger than the people actually buying its cars.

III. World Models & Physical AI

1. After K3.1 Pro, Kimi looks like it is moving toward embodied AI

• Summary: A TMTPost hands-on with Moonshot's (月之暗面) Kimi K3.1 Pro finds it far faster than K3, and reads the company's direction as edging into embodied AI.
• Source: TMTPost · 2026-10-25
• Editor's Take: Speed is table stakes now. The interesting claim is the pivot, because a lab shipping a fast chat model alongside a world-model roadmap is hedging both ways.

IV. Foundation Models

1. Kimi K3.1 Pro breaks into a global top four, first on agents

• Summary: Moonshot's (月之暗面) Kimi K3.1 Pro placed in the global top four and took first place among agent models in a new ranking, according to AI Era.
• Source: AI Era · 2026-10-25
• Editor's Take: Agent leaderboards reward tool use over raw reasoning. Useful signal, and a narrow enough slice to optimize against.

2. Anthropic's IPO shaves $500B off a $2T valuation

• Summary: At its investor day, Anthropic reported $16.8 billion in quarterly revenue, a second straight quarter of adjusted profit and its first positive free cash flow. The IPO still cut $500 billion from a valuation floated near $2 trillion, an exercise Huxiu calls a stress test.
• Source: Huxiu · 2026-10-25
• Editor's Take: Revenue growth and a valuation haircut in the same week is the market repricing AI risk, not losing faith in the product.

3. KRAFTON previews Raon-1 mini, a 30B-A4B challenger

• Summary: KRAFTON AI, the AI arm of the Korean game publisher, previewed Raon-1 mini, a 30B-A4B mixture-of-experts model trained from scratch with linear attention and speculative decoding. A full release and larger models are planned.
• Source: ITHome · 2026-10-25
• Editor's Take: A game studio training its own MoE is a cost calculation. Publisher-owned models only survive if the inference bill beats the license fee.

4. Alibaba's Qwen lead maps the personal-agent stack

• Summary: At China's computer congress on Oct 22, Jiang Guanjun, who heads AI algorithms for Alibaba's (阿里巴巴) Qwen unit, argued that a genuine personal agent needs models, memory and ecosystem moving together before it can offer proactive, personalized service.
• Source: ITHome · 2026-10-25
• Editor's Take: Memory is the honest answer. Without it, every personal agent resets to a stranger at the start of each session.

5. Block ships Dessa, a full-duplex voice model that runs locally

• Summary: Block, Jack Dorsey's company, released Dessa, a voice model that runs on-device and mimics laughter, pauses and the catch in a throat. It was trained entirely on human speech data using a single set of Nvidia GPUs.
• Source: ITHome · 2026-10-25
• Editor's Take: Local inference plus human-only training is a privacy pitch. The demo will settle whether it sounds like a person or a better phone tree.

6. Anthropic quietly trims Opus 5.5's reverse-engineering muscle

• Summary: Users who credited Claude Opus 5.5, shipped Sept 22, with decompiling and rebuilding games through vibe coding now report the capability has been quietly restricted.
• Source: ITHome · 2026-10-25
• Editor's Take: Ship a capability, then walk it back. The part worth noticing is that nobody announced the reversal.

7. Fei-Fei Li: language models are not the endgame

• Summary: In a long interview, Fei-Fei Li argues that large language models are not the endpoint of AI and that spatial intelligence is the next real frontier.
• Source: Huxiu · 2026-10-25
• Editor's Take: She has made this argument since ImageNet. It still holds, and the money is still flowing to tokens.

8. Eight AI safety declarations in a year, and no way to verify them

• Summary: On Oct 13, roughly 400 scientists, about 499 by publication, called on the UN to keep frontier models physically offline, stand up an international watchdog and force disclosure of serious incidents. The letter carries signatures from 34 countries; Huxiu audits what the pledges actually enforce.
• Source: Huxiu · 2026-10-25
• Editor's Take: Declarations are cheap and getting cheaper. The measured number is how many organizations publish a failed audit.

V. AI Software

1. CoreWeave is the AI trade's most exposed nerve

• Summary: CoreWeave fell another 10% to $55.60 on Oct 22, roughly 60% below its 52-week high. The same day, Nvidia's Jensen Huang called demand for neoclouds "off the charts," and CoreWeave is the largest of them.
• Source: Huxiu · 2026-10-25
• Editor's Take: Demand and share price pointing in opposite directions means the market is pricing the financing, not the customers.

2. Neolix raises nearly $400M as robot delivery goes to scale

• Summary: Neolix (新石器) closed a roughly $400 million Series E led by Boyu Capital and a national fund, joined by Alibaba's AEF NextGen Hong Kong fund, GGV and Temasek's Vertex Growth. Cumulative funding now tops ¥10 billion.
• Source: TMTPost · 2026-10-25
• Editor's Take: Unmanned delivery has left the demo phase for fleet economics. Round sizes like this say the winners have already been chosen.

3. China's first AI-made theatrical feature clears ¥1.16M

• Summary: "Sanxingdui: Future Past", billed as China's first AI-produced cinema feature, has grossed ¥1.162 million (~$160,000), according to Maoyan data.
• Source: ITHome · 2026-10-25
• Editor's Take: A niche film can clear a million yuan. AI does not change what a cinema audience is willing to pay for.

4. Alibaba's Qwen Go! lets users build and share apps by asking

• Summary: Alibaba (阿里巴巴) added a "My Apps" feature to its Qwen Go! agent, letting users assemble apps for work, study and daily life without writing code. The results can be shared and keep refreshing their data.
• Source: ITHome · 2026-10-25
• Editor's Take: Shareable, self-updating apps are the retention hook. The real product is distributing other people's creations.

5. AI agents reach the banks, and the banks start drafting disaster plans

• Summary: Pessimism about runaway AI has moved from tech billionaires to the large US banks that steer global capital, which are now preparing contingency plans, ITHome reports.
• Source: ITHome · 2026-10-25
• Editor's Take: Banks writing AI failure playbooks is a leading indicator. Institutions plan for what they expect to actually happen.

6. Subaru wants simulation to replace most crash tests

• Summary: Subaru plans a software-defined vehicle with over-the-air updates by the late 2020s and aims to cut next-generation development time by 40%, partly by replacing most physical crash tests with digital simulation. The company cites Nikkei Asia and NHK.
• Source: ITHome · 2026-10-25
• Editor's Take: Simulation shortens the routine work and leaves the expensive edge cases. Regulators will decide how much of the crash test moves into software.

7. Two Chinese AI-companion apps pulled from overseas iOS

• Summary: PolyBuzz and Tipsy Chat, both built by Chinese teams and reported to gross over ¥100 million a month, were pulled from overseas iOS stores. Huxiu ties the removals to Apple's rules on adult-adjacent content, with Zuoyebang behind one of them.
• Source: Huxiu · 2026-10-25
• Editor's Take: Companion apps are a revenue machine and a compliance hazard at once. Apple's store policy, not the model, decides how long that lasts.

8. A new benchmark asks whether AI can build an app from screenshots

• Summary: A benchmark out of HKUST (Guangzhou) tests whether models can reproduce an app's interaction from screenshots alone, per AI Era.
• Source: AI Era · 2026-10-25
• Editor's Take: Screenshot-to-app is the honest test for UI agents. It strips out the clean APIs most demos quietly lean on.

---

物界前沿 | EastVoice