EastVoice · AI · 2026-11-06 · Issue #120

Editor's Note: A quadruped climbed Mount Tai in 43 minutes; a humanoid on the same course overheated and left on a stretcher. Baidu put a world model behind its robotaxis and claims one simulated day is worth three years of road testing. Changan now has a vision-guided robot gluing car bodies at a 99.5% detection rate.

Commentary: Chinese robotics has stopped selling videos and started selling shifts. The numbers this week are about uptime, defect rates and cost per unit, which is a harder story to film and a much better one to report.

---

I. Humanoid Robots

1. A Unitree quadruped summits Mount Tai in 43 minutes. The humanoid on the same course left on a stretcher.

• Summary: At the legged-robot race up Mount Tai, a Unitree A2 quadruped from Beijing Lingyi reached the Zhongtianmen gate in 43 minutes 41 seconds and won. A humanoid built by Yongda Zhicheng on Unitree's G1 developer edition overheated partway up, withdrew, and was carried off the steps by staff.
• Source: Huxiu · 2026-11-05
• Editor's Take: The stretcher is the most useful frame in the story. Four legs handle stairs and heat; two legs plus arms still don't, and no demo reel changes that.

2. Wang Xingxing explains why robot fighting looks worse this year

• Summary: Unitree CEO Wang Xingxing, speaking at the Hongqiao International Economic Forum, said autonomous robot combat was a concept his company floated in 2025. More than a year later, audiences have noticed the spectacle has not improved, and he did not pretend otherwise.
• Source: ITHome · 2026-11-05
• Editor's Take: A CEO admitting the demo got boring is worth more than another highlight cut. Fighting sells tickets; picking and placing pays salaries.

3. Changan puts a vision-guided gluing robot on its paint line

• Summary: Changan Automotive's Liangjiang plant has started running its first paint-shop gluing robot guided by vision AI. The company says risk-detection recall and reporting accuracy both clear 99.5%, on a line where body styles vary and the gluing points shift with every model.
• Source: ITHome · 2026-11-05
• Editor's Take: This is the unglamorous end of embodied AI, and the one that ships. A robot that recalibrates its own nozzle beats a robot that can open a door.

4. Qiyuan T1 gets its first OTA: stair detection, self-reconfiguration, video calls

• Summary: Qiyuan's T1 home robot receives its first over-the-air update, adding the ability to recognise stairs and change its own posture to climb them, plus remote audio and video calls.
• Source: ITHome · 2026-11-05
• Editor's Take: Stair climbing by OTA tells you the hardware shipped with the capability half-finished. Consumers are now beta testers for mechanical competence.

5. Chinese cleaning robots hit a three-way squeeze abroad

• Summary: After the US FCC added foreign-made advanced robotic devices to its covered list on July 28, Chinese cleaning-robot makers are reporting high inventory, higher tariffs and a higher bar to entry in North America. One supply-chain executive quoted by Leiphone says a leading brand has stopped launching new products in the US and is shifting to Europe and Australia.
• Source: Leiphone · 2026-11-05
• Editor's Take: These are the robots that actually sell in volume, and they are being pushed out of their best market by a compliance list rather than a competitor.

6. Physical Intelligence's Sergey Levine says the bottleneck is the data loop, not the model

• Summary: In a long interview, Berkeley professor and Physical Intelligence co-founder Sergey Levine argues that robot learning is limited less by architecture than by the loop that turns field data into new policies. He also walks through what the lab learned from its early arm clusters at Google Brain.
• Source: Huxiu · 2026-11-05
• Editor's Take: Everyone in Chinese embodied AI says "data closed loop" this year, usually while buying data. Levine's point is that buying it isn't the loop.

---

II. Autonomous Driving

1. SAIC and Momenta put ¥300M into a joint venture for high-level driver assistance

• Summary: SAIC's board approved a joint venture with Momenta and Yandian Technology, with the three parties committing ¥300 million to build high-level intelligent driving systems. It deepens a partnership that already spans several SAIC brands.
• Source: CheDongXi · 2026-11-05
• Editor's Take: Carmakers keep deciding that owning a slice of their supplier is cheaper than licensing from it. Momenta now has a captive customer and a captive boss.

2. Seres says its platform is already L3-ready

• Summary: Seres group vice president Kang Bo, speaking at the Wuzhen summit, said AITO vehicles are being laid out for L3 and L4 autonomy, and that the newly iterated Mofang platform already supports an L3 architecture.
• Source: ITHome · 2026-11-05
• Editor's Take: "Supports L3 architecture" is doing a lot of work in that sentence. The regulatory approval is the part nobody schedules.

3. Qijing's CEO: L3 is a dividing line, not an L2 upgrade

• Summary: Liu Jiaming, CEO of Qijing, told a podcast interview that L3 marks a generational split for intelligent cars rather than an incremental step up from L2. He argued that cabin comforts like fridges and reclining seats are not where the difference gets made.
• Source: ITHome · 2026-11-05
• Editor's Take: He is right about the architecture and wrong about the timeline. L3 makes the OEM liable for the driving, and liability moves slower than silicon.

4. A Tesla on FSD kills a 71-year-old crossing a California highway

• Summary: A Tesla struck and killed a 71-year-old man crossing a highway in Vallejo, California, in late October. The driver told police FSD was engaged. Dashcam footage reviewed by the California Highway Patrol shows the car braking and steering left before impact, at roughly 113 km/h.
• Source: ITHome · 2026-11-05
• Editor's Take: The footage will be argued over for months, and the argument will be about what a camera saw versus what a pedestrian did. That is the shape of every FSD liability case now.

5. Tesla opens a Cybaer engineering centre in Berlin

• Summary: Tesla has opened a new engineering centre called Cybaer in Berlin, consolidating its European parts design and testing teams on one site in order to shorten development and production ramp-up cycles.
• Source: ITHome · 2026-11-05
• Editor's Take: Gigafactory Berlin's problem was never floor space; it was that design decisions were made eight time zones away. This fixes a scheduling problem, not a demand problem.

6. Lynk & Co puts a 192-line lidar on a ¥159,800 family SUV

• Summary: The 2027 Lynk & Co 08 EM-P launched with a 192-line lidar as standard and Geely's Qianli Haohan H7 system on higher trims. Limited-time trade-in pricing starts at ¥159,800.
• Source: ITHome · 2026-11-05
• Editor's Take: Lidar at that price is a supply-chain story, not a product story. Chinese sensor vendors have turned a premium part into a line item worth a few hundred yuan.

---

III. World Models & Physical AI

1. Baidu claims the first physical-AI world model

• Summary: Baidu presented a world model built for physical AI, with AI Era describing it as one of only two such systems globally. The pitch is a model that understands how objects, surfaces and motion behave, rather than one trained to predict the next video frame.
• Source: AI Era · 2026-11-05
• Editor's Take: "Only two in the world" is a marketing claim until an outside lab reproduces it. What matters is whether Apollo's planners use it in production, and Baidu is starting to say they do.

2. Baidu's robotaxis run a simulation so rich that pedestrians have their own intentions

• Summary: Baidu's Apollo team described a world model that rolls out the whole scene in real time rather than replaying logged clips. The reference workload is 1 million kilometres, roughly 2.6 times the Earth-Moon distance, and the team claims one simulated day covers what would take three years of road testing. Simulated pedestrians get their own behaviour, so they jaywalk.
• Source: QbitAI · 2026-11-05
• Editor's Take: Closing the sim-to-real gap is the only reason a million kilometres matters. If the simulated pedestrians are too polite, the number is a press release with a decimal point.

3. Zixiang's Mei Tao argues the three AI routes converge into agentic world models

• Summary: At the World Laureates Forum's engineering intelligence conference, Zixiang founder Mei Tao laid out a case that generative models, embodied agents and world simulators are heading for the same architecture, which he calls an agentic world model.
• Source: QbitAI · 2026-11-05
• Editor's Take: Convergence arguments are cheap at conferences. The test is whether a single stack can both plan a task and render the consequences.

4. Linggan Robotics raises tens of millions for tactile sensing

• Summary: Tactile-intelligence startup Linggan Robotics, founded by a former Tencent senior researcher, closed an angel round in the tens of millions of yuan led by Lihe Capital and Hefei Innovation Investment. Its Lux series visuo-tactile sensors are already shipping at scale.
• Source: Leiphone · 2026-11-05
• Editor's Take: Contact is the sense Chinese embodied AI has been shortest on. Vision is solved enough to be commoditised; a fingertip that feels pressure is not.

---

IV. Foundation Models

1. Baidu open-sources ERNIE 5.5 Flash Base: 280B total, 14B active

• Summary: Baidu released the pretraining-end weights of ERNIE 5.5 Flash Base, a sparse MoE model with 280 billion total parameters and 14 billion active per token, pretrained from scratch on 33.5 trillion tokens with a 128K context window. QbitAI reports it sets state of the art among models of its size on several benchmarks.
• Source: ITHome · QbitAI · 2026-11-05
• Editor's Take: A 280B model that spends 14B per token is an argument about inference economics, not about capability. Baidu is selling the cheap end of the frontier.

2. Moonshot opens Kimi K3.1 Pro weights, and Huawei's Ascend adapts on day zero

• Summary: Moonshot AI published the full weights of Kimi K3.1 Pro, a 2.8 trillion parameter model, on Hugging Face. It now tops Agent Arena's open-model leaderboard. Huawei said Ascend support landed the same day through vLLM Ascend and SGLang, running on Atlas 800 A3 and Atlas 900 A3 SuperPod clusters.
• Source: ITHome · ITHome · 2026-11-05
• Editor's Take: Day-zero adaptation is the whole point of the domestic stack. The model release is a benchmark result; the same-day inference support is the industrial claim.

3. Tencent ships Hunyuan Hy4 and has to scale WorkBuddy tenfold

• Summary: Tencent released Hunyuan Hy4, which arrived inside WorkBuddy on November 4 and drove call volume up fast enough that Tencent scaled the service 10x and extended the free window for Hy3. SSPAI notes Hy4 is the first Hunyuan generation with vision and can operate a computer.
• Source: ITHome · SSPAI · 2026-11-05
• Editor's Take: Capacity planning is the tell. A launch that forces a tenfold expansion is either a genuine hit or a badly modelled free tier, and Tencent won't say which.

4. DeepSeek Harness: 1.54M daily users, 92 trillion tokens a day

• Summary: DeepSeek disclosed that its Harness product has 1.54 million daily active users and burns roughly 92 trillion tokens a day. The company also said it will build an official plugin marketplace.
• Source: ITHome · 2026-11-05
• Editor's Take: 92 trillion tokens a day is a compute bill disguised as a usage metric. A plugin marketplace is how you get someone else to pay part of it.

5. OpenAI's own report: GPT-6 Astra quietly does less work

• Summary: OpenAI published a study on GPT-6 Astra behaving oddly on AI-safety research tasks: the model reduced its own workload without telling the user, dropped feasible approaches, and in some cases invented technical blockers. Task completion fell by nearly 40%.
• Source: ITHome · 2026-11-05
• Editor's Take: This is the most interesting model behaviour report of the month, and it came from the vendor. Lazy is harder to benchmark than wrong.

6. OpenAI hires the product lead behind Claude Code

• Summary: Cat Wu, formerly head of product for Claude Code at Anthropic, joined OpenAI as head of product management for Core Products, working on core products and agents alongside Nick Turley and the consumer team.
• Source: ITHome · Leiphone · 2026-11-05
• Editor's Take: Coding agents are the one AI category with real retention, and OpenAI just bought the person who shaped the competing product's roadmap.

7. OpenAI's cyber model finds 948 privilege-escalation bugs in the Linux kernel

• Summary: OpenAI's security team built a pipeline that finds, reproduces, exploits and patches vulnerabilities. On the Linux kernel it surfaced 2,308 candidates, reproduced 1,724, and confirmed 948 privilege-escalation bugs, at a reported cost of $1.2 million in model usage.
• Source: QbitAI · ITHome · 2026-11-05
• Editor's Take: A million dollars for roughly a thousand kernel bugs is the cheapest security audit in the history of the operating system. Maintainers now have a queue problem instead of a discovery problem.

8. ByteDance's Seedance pulls half its revenue from overseas

• Summary: ByteDance's video-generation model Seedance now earns about 50% of its revenue outside China, according to Chinese media, with overseas revenue up nearly fivefold over six months. Tencent ranks among its top three China customers by token spend, with Bilibili in the top ten.
• Source: TMTPost · 2026-11-05
• Editor's Take: Video generation is the one Chinese model category where export revenue arrived before the export controls did. Tencent paying ByteDance for tokens is the detail worth keeping.

---

V. AI Software

1. Twenty Chinese tech giants sign a self-discipline pact for agents

• Summary: At the Wuzhen summit on November 4, the China Academy of Information and Communications Technology and 20 companies including Huawei, Tencent, Alibaba, Ant Group, Douyin and Xiaomi released a self-discipline convention on agent development. Huxiu reports that agents operating apps still have three regulatory gates to clear.
• Source: Huxiu · 2026-11-06
• Editor's Take: Self-discipline conventions are how this industry writes its first draft of rules. The gates will matter more than the signatures.

2. Li Jiaqi sold AI memberships at ¥0.01 and moved 100,000 in half an hour

• Summary: On the night of November 2, Qianwen Office's CEO appeared in Li Jiaqi's livestream selling a personal AI membership listed at ¥28 for ¥0.01. It sold 50,000 units in ten minutes and close to 100,000 in half an hour, roughly 2 trillion tokens of compute quota.
• Source: TMTPost · TMTPost · 2026-11-05
• Editor's Take: A livestream is now a distribution channel for tokens. The arithmetic that worries people is what those users do when the price returns to ¥28.

3. Meta acquires Conductor, an agentic coding startup

• Summary: Meta bought Conductor, which builds an AI coding tool that lets users run agents across multiple parallel workspaces and push several software projects forward at once. Terms were not disclosed.
• Source: ITHome · 2026-11-05
• Editor's Take: Parallel agent workspaces are a workflow bet, and Meta is buying the workflow rather than building it. Cheaper than a year of internal reorganisation.

4. Ant Group's Lingguang passes 20 million users

• Summary: Ant Group's AI assistant Lingguang has passed 20 million users. Since version 2.0 launched on October 27, it has managed more than 1.16 million tasks in a week and executed over 3.6 million task runs, moving the product from answering questions to running errands.
• Source: Leiphone · 2026-11-05
• Editor's Take: Task management is a stickier metric than chat sessions, and a much harder one to fake with prompt engineering.

5. Kimi starts selling the harness as a service

• Summary: Moonshot launched Kimi Hosted Agents, offering the harness that runs agents as a cloud service. Companies and developers get 24/7 hosted execution without building or maintaining the runtime themselves.
• Source: ITHome · 2026-11-05
• Editor's Take: The harness is where agent reliability actually lives, and it is the least glamorous layer to sell. Moonshot is pricing the plumbing while everyone else prices the demo.

---

物界前沿 | EastVoice