EastVoice · AI · 2026-11-04 · Issue #118

Editor's Note: Three numbers from Anthropic's IPO paperwork do not sit comfortably together. Investors want the valuation trimmed toward $1.5 trillion, the long-range spending plan runs to $895 billion, and net revenue retention is 900%. The same week, Meta is borrowing up to $25 billion, Oracle wants another $20 billion, and Brookfield is putting $27 billion into a data center next to an English power station.

Commentary: Borrowed money buys the GPUs, and the GPUs sell tokens. Nobody has shown the full ledger of what sits between those two lines yet.

---

I. Humanoid Robots

1. Xiaomi puts its humanoid through an eight-hour factory shift, then sends it home to tidy up

• Summary: Xiaomi says its humanoid held an eight-hour shift on a production line doing console side-cover work, and separately chained together long-horizon chores in a home. The company collected two Wuzhen awards the same week, an Outstanding Contribution nod for the group and a Leading Technology award for its embodied foundation model.
• Source: ITHome · 2026-11-03
• Editor's Take: Eight hours is the number that counts. Any rig can do five minutes on camera; a shift is where thermal throttling and re-grasps show up.

2. A Chinese logistics player wants the AI to run the warehouse, not just the robots

• Summary: QbitAI reports on a next-generation autonomous warehouse where a scheduling model takes over operations and the robots handle picking. The pitch is that the site gets better the longer it runs, because the AI layer absorbs exceptions instead of escalating them to a supervisor.
• Source: QbitAI · 2026-11-03
• Editor's Take: Warehousing is the one embodied market with paying customers and a clean ROI. Whoever owns the scheduling layer owns the margin.

3. AgiBot takes the Wuzhen outstanding contribution award

• Summary: AgiBot (智元) was cited for systematic technical work and commercial deployment in embodied AI at the World Internet Conference opening on Nov 3, where Premier Li Qiang gave the keynote.
• Source: Leiphone · 2026-11-03
• Editor's Take: Awards are a cheap signal. The tell is that the judging language now mentions deployment, not just demos.

4. Wang Xingxing says Unitree's next fight is the robot brain, and he wants partners

• Summary: Speaking at Wuzhen, the Unitree founder said the company's focus is shifting to the robot brain and that it will look for collaborators worldwide rather than build the whole stack alone. He named physical AI as the area he wants to break.
• Source: ITHome · 2026-11-03
• Editor's Take: Unitree won the hardware cost war. Brains are a different fight, and it is the one where a Chinese vendor most needs outside help.

5. The robot business is turning into a rental business

• Summary: A Huxiu piece argues the money has moved downstream. Plenty of outfits advertising test-engineer roles turn out to be leasing operations rather than manufacturers like AgiBot or Galbot, and rental is where utilisation and cash flow actually add up.
• Source: Huxiu · 2026-11-03
• Editor's Take: Rental solves the demand problem and creates a maintenance problem. Whoever absorbs the downtime ends up owning the customer.

6. Li Auto walks back the embodied-AI pitch and goes back to selling cars

• Summary: An investor's essay on Huxiu tracks a 180-degree strategy reversal at Li Auto inside six months, from premium embodied-AI framing to the things buyers already pay for.
• Source: Huxiu · 2026-11-03
• Editor's Take: Embodied AI as a marketing layer does not survive a price war. Li Auto's retreat is the clearest read yet on what consumers will fund.

---

II. Autonomous Driving

1. Tesla's robotaxi network adds Fort Worth, still with nobody in the driver's seat

• Summary: Service opened in Fort Worth at 6am Central on Nov 2, with no safety monitor aboard, matching Tesla's other Texas launches.
• Source: ITHome · 2026-11-03
• Editor's Take: Adding a city is easy when the geofence is generous. The real test is whether a storm-season Fort Worth behaves like Austin.

2. Japan moves to certify cars that have no steering wheel at all

• Summary: Nikkei Asia reports Tokyo is weighing a rewrite of vehicle safety certification so cars without steering wheels, brake pedals or accelerator pedals can be type-approved under one standard. The rule would cover L4 vehicles driving autonomously under set conditions as well as L5.
• Source: ITHome · 2026-11-03
• Editor's Take: Japan is writing the paperwork for robotaxis before it has the robotaxis. Regulation first is how a lagging market buys time.

3. Dongfeng ships its own ADAS alongside Huawei's in the same SUV line

• Summary: Dongfeng's Yipai brand added a lidar trim of the M8, listing at ¥179,800-185,800 and selling from ¥159,800 after incentives. It runs Dongfeng's in-house Tianyuan stack instead of the Huawei Qiankun system offered on other trims.
• Source: ITHome · 2026-11-03
• Editor's Take: A state-owned automaker fielding its own stack next to Huawei's is the quiet story. Supplier leverage in Chinese ADAS is not settled.

---

III. World Models & Physical AI

1. Baidu's Apollo World bets that physics beats pixels

• Summary: At AI Day in Shenzhen, Baidu showed Apollo World, an agent-driven omnimodal world model for physical AI, and proposed a "2+3" rubric for judging such models. The company's line is that fidelity to physical law and causality matters more than how good the render looks.
• Source: Leiphone · ITHome · 2026-11-03
• Editor's Take: Baidu is selling simulation as data. If the world model is faithful enough, road miles stop being the bottleneck, and that undercuts everyone's testing budget.

2. Ant Group's Lingbo brings an open-source robot brain to Wuzhen

• Summary: Ant's embodied unit showed a full-stack open-source general brain for robots, spanning model research, open collaboration and deployment. The argument is that open source is moving out of text and code and into the physical world.
• Source: Leiphone · 2026-11-03
• Editor's Take: Open weights are harder in robotics than in LLMs because the hardware varies. Ant is betting breadth of contributors beats one vendor's data.

3. Xiaomi tests its embodied base model on its own staff

• Summary: Xiaomi-Robotics-1 went into trial operation in the company's youth apartments, planning and executing desk tidying and trash clearing without step-by-step instruction.
• Source: ITHome · 2026-11-03
• Editor's Take: Staff housing is a clever way to collect messy real data without a customer. It is also a demo curated by people who want it to work.

4. MetaCircle raises two rounds in four months for a model that grows intuition

• Summary: MetaCircle (元环智能), founded four months ago by a Tsinghua AI assistant professor, closed two rounds totalling several hundred million yuan. The money funds its meta-model algorithms, early products, hiring and self-built benchmarks.
• Source: Leiphone · 2026-11-03
• Editor's Take: That speed tells you where Chinese capital thinks the next wedge is. Self-built benchmarks are the tell, they are building the scoreboard they intend to win.

---

IV. Foundation Models

1. MIT has GPT-6 Astra edit another model's weights, and math scores jump sixfold

• Summary: CSAIL researchers had GPT-6 Astra read and modify the weights of Qwen3.5-0.8B directly, touching about 0.016% of parameters. Math scores rose to roughly 6x the base model, a method they call weight-space programming.
• Source: ITHome · QbitAI · 2026-11-03
• Editor's Take: Editing weights instead of prompting is a different product category. It also needs read access to someone else's checkpoint, which is a licensing fight waiting to happen.

2. Half of Seedance's revenue now comes from outside China

• Summary: ByteDance's video model takes about 50% of revenue overseas, with that slice up nearly 5x in six months, according to a report cited by ITHome. Tencent and Bilibili are named as major domestic customers.
• Source: ITHome · 2026-11-03
• Editor's Take: Chinese video models are winning abroad on price and speed, not brand. Same playbook that worked for short-video apps.

3. Kimi plugs into Unity and Unreal, aiming at game prototyping

• Summary: Moonshot shipped a game-development and creative workflow built on its K3.1 models, Kimi Code, Kimi Work agents and plugins for Unity, Unreal and Blender. The target is validating gameplay ideas faster, not shipping final builds.
• Source: ITHome · 2026-11-03
• Editor's Take: Game engines are where Chinese model vendors can charge real money. Studios there pay per seat and tolerate lock-in.

4. SenseTime's SenseNova 6.8 Flash leans on memory and design output

• Summary: The production release adds long-term memory and cross-session task handoff, with design coverage stretching from animated slides and posters to web pages, SVG and Blender 3D work.
• Source: ITHome · 2026-11-03
• Editor's Take: Office and design generation is a crowded lane and SenseTime is late. Memory across sessions is the only piece here that is not commodity.

5. Joe Tsai puts Alibaba's open-model count at more than 470

• Summary: Speaking at the Wuzhen opening, Alibaba's chairman said compute demand keeps outrunning supply and argued that only open source gets AI to mass adoption. He said the Qwen family has passed 470 open releases.
• Source: ITHome · 2026-11-03
• Editor's Take: Alibaba has turned open weights into distribution. Every Qwen fork is a customer who eventually buys cloud.

6. Anthropic's net revenue retention hits 900%

• Summary: As of Sept 30, net revenue retention was about 900%, meaning the same accounts spent roughly nine times more than a year earlier. High-spend customers passed 11,900.
• Source: ITHome · 2026-11-03
• Editor's Take: That number is what happens when accounts move from pilots to production. It cannot repeat, which is exactly why the IPO math is being argued over.

7. Huawei Cloud retires two Pangu models with three weeks' notice

• Summary: OpenPangu-2.0-Flash and Pro stop selling on Nov 6 and go offline Nov 30, per a Nov 1 notice. New Pangu models are said to be in planning.
• Source: ITHome · 2026-11-03
• Editor's Take: Retiring a model mid-quarter is a service-promise problem. Enterprise buyers remember migration pain longer than benchmark scores.

8. DeepSeek open-sources a super-agent prototype under MIT

• Summary: The Harness team published a design note on more autonomous agents on Oct 31 and put dsh-superagent-prototype on GitHub under MIT. The write-up is framed as engineering practice rather than research.
• Source: TMTPost · 2026-11-03
• Editor's Take: DeepSeek's edge has been cost per token, and agents burn tokens by design. This is the company defending its own moat.

---

V. AI Software

1. Apple Intelligence finally lands in China, with Qwen doing part of the work

• Summary: The first iOS 27.3 developer preview brings Apple Intelligence and Siri to mainland handsets, iPhone 15 Pro and newer. The build adds AirPods live translation and lets the iPhone hand a wider set of tasks to Alibaba's Qwen.
• Source: ITHome · 2026-11-04
• Editor's Take: A China build with a local model partner is the template for every market Cupertino cannot serve alone. Watch what stays out of this build.

2. Siri handles 2.5 billion requests a day, and some of it will be paid

• Summary: On the earnings call, CEO John Ternus repeated the 2.5 billion daily request figure and confirmed that some advanced AI features will carry a charge.
• Source: ITHome · 2026-11-03
• Editor's Take: 2.5 billion is an installed-base number, not a demand number. Charging for the top tier will show how much of that base wants the new Siri at all.

3. Goldman Sachs pulls advanced coding AI from most of the floor

• Summary: Bloomberg reports the bank barred most staff from advanced coding-focused AI tools after weighing productivity against risk, and that the decision annoyed employees.
• Source: ITHome · 2026-11-03
• Editor's Take: Banks are where agent risk gets a price first. Compliance wins there until someone can audit an agent's actions line by line.

4. Two Chinese personal agents, five days, one Meta Muse to beat

• Summary: A hands-on comparison on Huxiu pits Lingguang and Today AI against each other and against Meta's Muse. Both are chasing an assistant that acts across apps rather than one that chats.
• Source: Huxiu · 2026-11-03
• Editor's Take: Muse has a year of Western usage data behind it. The Chinese clones are shipping features faster than they are collecting feedback.

5. After close to $100M, a Chinese startup bets on an "action partner" for messy environments

• Summary: GeekPark profiles a Chinese embodied-AI company that raised close to $100 million and built its product for open, dynamic real-world settings rather than fixed factory cells, pitching it as handing daily chores back to people.
• Source: GeekPark · 2026-11-03
• Editor's Take: "Open dynamic world" is the hardest claim in robotics. Most demos quietly retreat to a structured room once the camera is off.

6. Hong Kong University writes the first Know Your Agent report

• Summary: The KYA research group released a report at Hong Kong Fintech Week covering identity, authorisation and liability once agents take part in real economic activity, with nine categories of enterprise practice included.
• Source: Leiphone · 2026-11-03
• Editor's Take: KYA is the layer agents need before banks let them touch money. Whoever writes the standard shapes the market that follows.

7. The consumer AI half-life is down to 76 days

• Summary: A Leiphone analysis of the consumer AI app market finds products dying in as little as 76 days, with even apps holding tens of millions of users going under.
• Source: Leiphone · 2026-11-03
• Editor's Take: Distribution is not retention. A 76-day half-life means the category is still being discovered, not competed in.

---

物界前沿 | EastVoice