EastVoice · AI · 2026-11-02 · Issue #116

Editor's Note: Unitree reported its first quarter as a public company and the shape of it is instructive — more robots sold, thinner profit on each yuan of revenue. The same week, its stock ran 60% in ten sessions and the embodied-AI funding count nearly halved month on month. Meanwhile DeepSeek published the architecture behind its own agent stack and put a prototype on GitHub.

Commentary: Chinese robotics has spent two years proving it can build the body. The next bill is for the model and the compute that make the body useful, and this quarter is the first time that bill showed up in a public filing.

---

I. Humanoid Robots

1. China puts humanoids on 3,100 steps of Mount Tai

• Summary: The CMG World Robot Climbing Competition opens on Nov 2 at the foot of Mount Tai in Tai'an, Shandong. Machines must climb the ancient stone stairway from Hongmen to Zhongtianmen — 3,100 steps, 600 meters of vertical gain — with broadcast coverage of how each one sees the route, picks a line and improvises when the line fails.
• Source: ITHome · 2026-11-01
• Editor's Take: Stairs are the cruellest test in legged robotics. No two steps are the same height, there are no lane markings, and there is no reset button. Finishing does not mean close to useful, but the failures will be worth more than the finish.

2. Convenience-store robots: ten thousand lab reps, then a dropped bottle at the shelf

• Summary: QbitAI's field report from retail deployments — machines selling drinks and ice cream — finds robots that survived thousands of trials in the lab still fumble bottles in front of a real shelf, where lighting, stock and customers never repeat the same way twice.
• Source: QbitAI · 2026-11-01
• Editor's Take: A demo video measures the best run. A store measures the worst run, thirty times a day, with a customer standing there watching.

3. WeChat Pay's service robot clocks in at a 7-Eleven

• Summary: On Oct 29, Guangdong's 2,000th 7-Eleven opened inside Tencent's Shenzhen campus with a WeChat Pay-branded service robot working the floor. Built on Tencent Hunyuan, it greets customers, recommends products and bundles, and takes the order; the human still scans and pays.
• Source: ITHome · 2026-11-01
• Editor's Take: Greeter first, cashier never. Retail deploys robots where a mistake costs a shrug, and that placement says more about readiness than any benchmark score.

4. Unitree's first post-IPO quarter: more robots sold, less profit per yuan

• Summary: Unitree reported Q3 on Oct 28, its first since listing. Revenue rose while profit per unit of revenue fell. The company is now funding models and compute out of hardware margin — the bill every embodied-AI firm is beginning to receive.
• Source: Huxiu · 2026-11-01
• Editor's Take: Hardware margin was always going to carry the model team for a while. The open question is whether the model team ever hands that margin back.

5. Unitree's stock runs 60% in ten sessions on an unchanged P&L

• Summary: After a deep drawdown following listing, Unitree rallied more than 60% across the ten trading days from Oct 15, taking its market capitalisation back above ¥250 billion. The move rests on expectations for embodied AI at scale rather than on the quarter just reported.
• Source: Huxiu · 2026-11-01
• Editor's Take: A 60% move on unchanged fundamentals is a vote about the category, not the company. When the category cools, the multiple leaves first and the orders leave second.

6. Embodied-AI money is rotating into robot skin

• Summary: IT Juzi counted 33 embodied-AI deals worth ¥10.1 billion in October, against 60 deals worth ¥11 billion in September. Capital is shifting from whole-body hardware toward components — tactile skin, joints, the parts that decide whether a hand can actually grip.
• Source: TMTPost · 2026-11-01
• Editor's Take: Deal count halving while total value barely moves means bigger and fewer cheques. That is what a market looks like once it starts picking winners instead of funding categories.

7. Primary-market investors have gone quiet on physical AI

• Summary: Cailianpress surveyed funds that had backed robotics and embodied AI. Jensen Huang and Unitree's Wang Xingxing keep talking the sector up, but several investors told the paper they have stopped writing new cheques and are waiting for deployment data before the next round.
• Source: Huxiu · 2026-11-01
• Editor's Take: The hype cycle and the funding cycle run on different clocks. Founders who closed in the first half have runway; the ones raising now will hear the word "milestone" a great deal.

8. Meta taps China's Fourier for a humanoid prototype

• Summary: The Information reported that Meta built a humanoid prototype with Chinese robotics firm Fourier, described as a physical embodiment of its Muse model. Neither the hardware specification nor a production plan has been disclosed.
• Source: ITHome · 2026-11-01
• Editor's Take: Meta's gap is not the model, it is the actuator and the supply chain underneath it. Partnering with a Chinese vendor buys two years and hands that vendor the learning.

---

II. Autonomous Driving

1. Geely sells 310,200 cars in October, up 1%

• Summary: Geely's October total came to 310,200 units, up about 1% year on year. The first ten months reached 2.5457 million, up 3%. Geely brand moved 255,700 (+4%) and Galaxy 153,400 (+20%).
• Source: ITHome · 2026-11-01
• Editor's Take: Flat headline, strong sub-brand. The premium line is doing the work while the base brand spends its time defending price.

2. HIMA passes 1.6 million cumulative deliveries at ¥400,000 average

• Summary: Huawei's HIMA delivered 33,537 vehicles in October at an average transaction price above ¥400,000, taking cumulative deliveries past 1.6 million — the fastest any Chinese new-energy brand has reached that mark.
• Source: ITHome · 2026-11-01
• Editor's Take: The ¥400,000 average is the number that matters. Huawei is selling software-defined cars at German-brand prices and buyers are paying it for the ADS badge.

3. Huawei's Qiankun passes 17.3 billion km of assisted driving

• Summary: Huawei's Intelligent Automotive Solution unit said cumulative assisted-driving distance on Qiankun passed 17.3 billion km, and published a compilation of emergency interventions as evidence of edge-case handling.
• Source: ITHome · 2026-11-01
• Editor's Take: Mileage is the cheapest metric to quote and the hardest to audit. Disengagement rate per thousand kilometres, in the cities where the system is actually switched on, is the number worth arguing about.

4. Hongqi's new H9 ships Huawei ADS 5 with road preview

• Summary: FAW's Hongqi previewed the new H9 sedan: closed-loop air suspension and iDCS adaptive dampers as standard, plus road preview that reads the surface ahead. The driver-assistance stack is Huawei ADS 5.
• Source: ITHome · 2026-11-01
• Editor's Take: Road preview is a chassis feature dressed as an AI feature. It works, and it is exactly the sort of thing a buyer notices inside the first five minutes.

5. Xiaomi Auto clears 50,000 deliveries in October

• Summary: Xiaomi delivered more than 50,000 vehicles in October and put a limited run of N70 inventory on sale with three-year interest-free financing.
• Source: ITHome · 2026-11-01
• Editor's Take: Fifty thousand a month puts Xiaomi in the top tier of Chinese EV makers two years after shipping its first car. From here the constraint is service capacity, not order book.

6. L4 needs bigger models, which means different chips

• Summary: TMTPost argues that scaling to L4 pushes driving models past what general-purpose automotive SoCs can serve. Memory bandwidth, on-chip interconnect and thermal budget have to be designed around the model, rather than shrinking the model to fit the silicon.
• Source: TMTPost · 2026-11-01
• Editor's Take: Every autonomy programme eventually becomes a chip programme. The teams that admit it early spend less on re-spins and lose less time to the admission.

---

III. World Models & Physical AI

1. Westlake University opens a world-model research centre

• Summary: Westlake University opened its World Model Research Center at the Yungu campus on Oct 31, three years after it began assembling the team. The centre bets that the step beyond large language models is a model of how physical events unfold.
• Source: QbitAI · 2026-11-01
• Editor's Take: A university centre is a five-year bet placed with ten-year money. That is the right horizon for world models and the wrong one for anyone who needs a product next quarter.

2. Tsinghua's Andrew Yao wants AI to reason like Newton

• Summary: A team led by Andrew Yao and Yuan Yang at Tsinghua published work on an Algebraic World Model, representing physical regularities as algebraic structure a model can manipulate rather than as pixels or tokens to predict.
• Source: QbitAI · 2026-11-01
• Editor's Take: Predicting the next frame gets you a good video generator. Finding the invariant that makes the next frame inevitable gets you something a robot can plan with.

---

IV. Foundation Models

1. OpenAI, Google, Microsoft and Anthropic agree on a crawler opt-out

• Summary: The four companies will honour a shared standard that lets publishers refuse AI crawling in a single declaration, instead of negotiating with each vendor separately. The mechanism is designed to apply across multiple AI operators at once.
• Source: ITHome · 2026-11-01
• Editor's Take: A standard everyone adopts voluntarily is a standard everyone can ignore voluntarily. It does move the argument from "did you ask" to "did you respect the signal", which is progress of a kind.

2. DeepSeek open-sources the architecture behind its super agent

• Summary: On the evening of Oct 31, DeepSeek's Harness team published the design for a Super-agent system and released a research prototype on GitHub, aimed at agents that take initiative on day-to-day work rather than waiting for instructions. The first public build targets a single repository.
• Source: Huxiu · 2026-11-01
• Editor's Take: Publishing the architecture is a recruiting move as much as a technical one. Whoever defines how agents split work and hand it off defines the toolchain everyone else ends up building on.

3. China's inference volume passes training for the first time

• Summary: CCID's industrial-trends report, cited by CCTV, found that China's AI inference volume passed training volume for the first time, with daily token calls above 175 trillion as of June. The shift marks the move from model R&D into industrial deployment.
• Source: ITHome · 2026-11-01
• Editor's Take: Inference overtaking training is the line where AI stops being a research budget and becomes an operating cost. That changes who signs the cheque every month.

4. SenseTime open-sources a 9B model it says matches 27B

• Summary: SenseTime released SenseNova Loop-9B-Preview on Oct 30, built on a Looped Transformer that reuses layers to raise effective computational depth. The company claims it holds up against models three times its size on complex tasks.
• Source: ITHome · 2026-11-01
• Editor's Take: Looped architectures trade parameters for latency. A good deal if you are serving thousands of requests, a bad one if your user is sitting there waiting.

5. Pika ships H3 Flex on MiniMax's H3, picture and sound in six steps

• Summary: Pika released H3 Flex, a derivative of MiniMax H3, with adapter weights and inference code open. Six denoising steps generate video and audio together; inputs can be text, first and last frames, or multimodal references.
• Source: Leiphone · 2026-11-01
• Editor's Take: Joint audio-video in six steps is a serving-cost argument, and serving cost is what decides whether generated video becomes a product or stays a demo reel.

6. Mathematicians push back: you distilled the world, and won't let us distill you

• Summary: After Terence Tao's public criticism of OpenAI, Fields Medalist Deng Yu added his own, arguing that labs trained on the whole world's output have no standing to forbid others from training on theirs. Chinese mathematicians are broadly unconvinced that solved problems mean solved mathematics.
• Source: Huxiu · 2026-11-02
• Editor's Take: The complaint is about asymmetry rather than copyright. Every lab wants open data coming in and closed weights going out, and the ones that say so loudest are usually the ones with the most weights.

7. Agents are rewriting the server CPU bill

• Summary: TrendForce expects server CPU shipments to grow 31% in 2026 as AI-agent workloads spread, with further growth in 2027 as inference and agentic services scale. The CPU stops being the supporting act inside an AI server.
• Source: TMTPost · 2026-11-01
• Editor's Take: Agents spend most of their time on tool calls, state and orchestration, and that work lands on CPU rather than tensor cores. The accelerator-only buildout was always going to need a second budget line.

8. ChatGPT and Codex start sharing one quota in December

• Summary: OpenAI said that from December, as Work's agent capabilities fold into the main ChatGPT experience, Plus, Pro and Business subscribers will draw on a single allowance across ChatGPT and Codex.
• Source: ITHome · 2026-11-01
• Editor's Take: Unlimited chatting was a customer-acquisition cost, and acquisition costs end. The interesting question is what heavy Codex users do in January, when the meter is shared and visible.

---

V. AI Software

1. Ant Group's Lingguang 2.0: an assistant that still can't leave the phone

• Summary: Geekpark tested Lingguang 2.0, Ant Group's personal AI assistant. It handles errands and cross-app tasks well inside its lane and hits the same wall Doubao's phone assistant does — it can plan the action but cannot always execute it where the app actually lives.
• Source: Geekpark · 2026-11-01
• Editor's Take: The bottleneck moved off the model and onto the operating system's permission model. No company shipping an assistant can fix that on its own.

2. No code, still shipping: the founders who build it themselves

• Summary: Huxiu profiles founders who took development in-house. One wanted an enterprise AI service, was quoted $105,000 for the build, shipped it on Replit for under $400, and landed a first client worth $15,000 within five days.
• Source: Huxiu · 2026-11-02
• Editor's Take: The $400 build is real. So is the maintenance bill nobody quotes in week one. Agencies will survive on the second year of the contract, not on the first.

3. Double 11 meets the AI shopper, and the arithmetic stops adding up

• Summary: Huxiu and partners ran one shopping request through Taobao's built-in AI, JD's AI shopping assistant, Doubao and SMZDM's assistant. The four returned different prices, different bundles and different readings of what the promotion actually entitled the buyer to.
• Source: Huxiu · 2026-11-02
• Editor's Take: Promotions were designed to be legible to humans and opaque to scripts. An agent that reads every rule is the first shopper with a real chance of beating the house.

4. A ¥20,000 quote negotiated to ¥2,000: the FDE who went back to a job

• Summary: A forward-deployed engineer told Huxiu he wound down his own practice after a client countered his ¥20,000 quote with ¥2,000. The work was deliverable; the client's price expectation had been set by tools that let them try it themselves first.
• Source: Huxiu · 2026-11-01
• Editor's Take: Once the buyer can prototype in an afternoon, the seller is pricing labour rather than outcomes. Forward deployment only pays where the integration is genuinely hard.

5. A design leaderboard where several models score zero

• Summary: A new leaderboard evaluates models on real engineering-design tasks rather than exam-style problems. Several entries came back with zero.
• Source: AI Era · 2026-11-01
• Editor's Take: A model that can describe a design is not the same as one that produces a design which holds. Text benchmarks structurally cannot see the difference.

6. When agents work in parallel, someone still has to hold the human's attention

• Summary: Huxiu's essay on DeepSeek's super-agent design argues the scarce resource is not compute but attention. Once several agents run at once, the human becomes the queue, and the system's real job is deciding which interruption has earned the right to happen.
• Source: Huxiu · 2026-11-01
• Editor's Take: Every agent product eventually ships a notification policy. Most ship a bad one, then call it a feature.

---

Physical World Frontier | EastVoice