Headlines · Hot List · Forum · 中文

EastVoice

PHYSIX FRONTIER EASTVOICE

2026.11.03 · Tuesday

Issue 073

FIVE MINUTES A DAY, THE GLOBAL AI SHIFT

90 teams

Teams in CMG World Robot Mount Tai Race climb

$17 billion

Anthropic Q3 revenue, roughly 11x year-earlier

1.6 trillion

Meituan LongCat-2.5 total parameters, domestic silicon training

2 trillion tokens

Li Jiaqi livestream token usage sold in 30 minutes

Top Stories 3 · dual commentary | Section Picks 21 · one-line takes

TOP STORIES

Unitree takes one gold and two silvers at China's first robot mountain race; Anthropic's Q3 revenue reaches $17 billion; Meituan's LongCat-2.5: 1.6 trillion parameters, trained entirely on Chinese silicon.

1Unitree takes one gold and two silvers at China's first robot mountain race

The CMG World Robot Mount Tai Race sent roughly 90 teams up the 3.9-kilometer, 3,100-step climb from Red Gate to Zhongtianmen on November 2. Unitree's Beijing Lingyi crew won the legged-robot sprint and added two second places, with the course's 600-meter elevation gain doing most of the sorting.

What it means for you|Unitree's Beijing Lingyi crew won the legged-robot sprint and added two second places, with the course's 600-meter elevation gain doing most of the sorting.

Editor's take|Stone steps are a brutal filter. Flat-ground demos don't survive 3,100 consecutive edges.

Source: ITHome (2026-11-02)

2Unitree's G1 overheated on the mountain and left on a stretcher

A Unitree G1 withdrew from the same race after overheating, and footage showed it being carried down the mountain on a stretcher. Robots were released at roughly one-minute intervals, and heat buildup on continuous stair climbing ended the run early.

What it means for you|A Unitree G1 withdrew from the same race after overheating, and footage showed it being carried down the mountain on a stretcher.

Editor's take|The stretcher is the most useful data point of the weekend. Thermal budget, not gait, is what stops a humanoid on a real hill.

Source: ITHome (2026-11-02)

3Meta reportedly built a humanoid with Fourier in three months; Fourier won't say

Reports say Meta worked with Chinese robotics firm Fourier to produce a prototype called Max in about three months, standing roughly 1.5 meters and weighing 48.1 kilograms, with an Nvidia Thor chip inside. Asked to confirm, Fourier told First Financial it was "not convenient to reply."

What it means for you|Reports say Meta worked with Chinese robotics firm Fourier to produce a prototype called Max in about three months; Fourier told First Financial it was "not convenient to reply."

Editor's take|"Not convenient to reply" is a confirmation with better lawyers. The three-month timeline says more about Chinese supply chains than about Meta's hardware team.

Source: ITHome (2026-11-02)

SECTION PICKS

21 AI stories worth your time, each with a one-line take.

Humanoid RobotsUBTech ships Thinker-Sim, a simulation stack built around grading robots

UBTech released Thinker-Sim on November 2, a simulation platform that covers digitizing real scenes, generating training data at scale, and benchmarking robot capability in one pipeline. The pitch is turning a physical site into a test rig quickly.

What it means for you|UBTech released Thinker-Sim on November 2, a simulation platform covering digitizing real scenes, generating training data at scale, and benchmarking robot capability in one pipeline.

One-line take|Every humanoid company now sells a simulator. The ones that matter will be judged on whether their sim score predicts field performance.

Source: ITHome (2026-11-02)

Humanoid RobotsFaraday Future's robot arm reports 351 units in October, past 1,000 cumulative

Jia Yueting's weekly update put FFR robot sales and shipments at 351 units for October, bringing the cumulative total to 1,168 and past the 1,000 mark for the first time. The company is guiding for more than 400 units a month in November and December and says deliveries are gross-margin positive.

What it means for you|Jia Yueting's weekly update put FFR robot sales and shipments at 351 units for October, bringing the cumulative total to 1,168 and past the 1,000 mark for the first time.

One-line take|Four hundred robots a month is a supply-chain achievement. Whether those units do paid work is a separate question nobody in the announcement answers.

Source: ITHome (2026-11-02)

Humanoid RobotsApple's HomeView Ultra robot hardware was frozen two years ago, and Apple still argues about it

The Information's Aaron Tilley reports that Apple's robot version of HomeView, variously called HomeView Ultra or HomeView Max, has had finalized hardware for about two years while internal disagreements over the product continue. The prototype pairs a display with a moving base.

What it means for you|The Information's Aaron Tilley reports that Apple's robot version of HomeView has had finalized hardware for about two years while internal disagreements over the product continue.

One-line take|Two years of frozen hardware and no launch is what happens when the demo is easy and the job to be done is unclear.

Source: ITHome (2026-11-02)

Autonomous DrivingA Starship delivery robot ran a red light in Tallinn and nine bus passengers were hurt

On the night of October 26, a Starship delivery robot ran a red light in central Tallinn, forcing a bus to brake hard; nine passengers were injured. Starship blamed an advertising billboard for confusing the robot and denied leaving the scene.

What it means for you|On the night of October 26, a Starship delivery robot ran a red light in central Tallinn, forcing a bus to brake hard; nine passengers were injured.

One-line take|A sidewalk robot with a top speed near walking pace still found a way to hurt nine people. The billboard excuse is an admission that the perception stack was reading the street wrong.

Source: ITHome (2026-11-02)

Autonomous DrivingXPeng's MONA L03 passes 20,000 October deliveries, third straight monthly gain

XPeng said the MONA L03 delivered more than 20,000 units in October, its third consecutive month of growth. The model remains the volume floor under XPeng's driver-assist rollout.

What it means for you|XPeng said the MONA L03 delivered more than 20,000 units in October, its third consecutive month of growth.

One-line take|Cheap cars carrying good software is the only combination that gets urban ADAS data at Chinese scale.

Source: ITHome (2026-11-02)

Autonomous DrivingAITO says the new M9 lineup passed 50,000 deliveries in 20 weeks

Huawei's AITO brand reported 50,000 cumulative deliveries for the new M9 series in 20 weeks, with the long-wheelbase M9 Ultimate passing 5,000 in its first 50 days. AITO also published an owner profile claiming nearly 70% of Ultimate buyers live in households with after-tax income above ¥1 million (~$140,000).

What it means for you|Huawei's AITO brand reported 50,000 cumulative deliveries for the new M9 series in 20 weeks, with the long-wheelbase M9 Ultimate passing 5,000 in its first 50 days.

One-line take|Publishing your buyers' income bracket is a pricing strategy dressed as transparency.

Source: ITHome (2026-11-02)

World Models & Physical AIYann LeCun tells Dartmouth that generative AI is a dead end

Speaking at Dartmouth, where the 1956 workshop that named the field took place, Turing laureate Yann LeCun argued that autoregressive generative models cannot get to machine intelligence and that world models are the path forward. He has made the case before; this time he made it in the room where the field started.

What it means for you|Speaking at Dartmouth, where the 1956 workshop that named the field took place, Turing laureate Yann LeCun argued that autoregressive generative models cannot get to machine intelligence and that world models are the path forward.

One-line take|LeCun has been saying this since before it was contrarian. The part worth watching is who is now funding world models while still shipping chatbots.

Source: Huxiu (2026-11-02)

World Models & Physical AIWhere the strongest intelligence actually goes

A Huxiu essay in the "after all AI obey" series opens with an elderly woman steadying herself on a bedside robot and argues that the destination of frontier capability is not a leaderboard position. The measure it proposes is whether a machine can help someone stay in their own home.

What it means for you|A Huxiu essay opens with an elderly woman steadying herself on a bedside robot and argues that the destination of frontier capability is not a leaderboard position.

One-line take|Eldercare is the hardest embodied benchmark there is: soft contact, unpredictable humans, and no reset button.

Source: Huxiu (2026-11-02)

World Models & Physical AIHow embodied agents build a 3D picture of a room

An AI Era piece examines the layer between perception and action, tracking how observation agents assemble three-dimensional spatial understanding when they enter an unfamiliar room. The argument is that spatial reasoning, not object detection, is what lets an agent decide where to look next.

What it means for you|An AI Era piece tracks how observation agents assemble three-dimensional spatial understanding when they enter an unfamiliar room, arguing spatial reasoning lets an agent decide where to look next.

One-line take|Active looking is the missing piece in most home-robot demos. They see everything at once and understand nothing about where to spend attention.

Source: AI Era (2026-11-02)

Foundation ModelsAnthropic's Q3 revenue reaches $17 billion, up roughly 11x

Anthropic disclosed in an amended S-1 that revenue for the quarter ended September 30 came in near $17 billion (about ¥113.2 billion), roughly 11 times the year-earlier figure. The filing lands ahead of what bankers expect to be a record listing.

What it means for you|Anthropic disclosed in an amended S-1 that revenue for the quarter ended September 30 came in near $17 billion, roughly 11 times the year-earlier figure.

One-line take|Eleven-fold growth at that base is not a startup curve. It's a compute-conversion business, and the compute bill arrives later in the same document.

Source: ITHome (2026-11-02)

Foundation ModelsMeituan's LongCat-2.5: 1.6 trillion parameters, trained entirely on Chinese silicon

Meituan released LongCat-2.5 on November 2, a natively multimodal model with 1.6 trillion total parameters aimed at long-horizon agent tasks. The company says the full training run stayed on domestic accelerators, making it the largest training job completed on Chinese compute to date.

What it means for you|Meituan released LongCat-2.5 on November 2, a natively multimodal model with 1.6 trillion total parameters aimed at long-horizon agent tasks, trained on domestic accelerators.

One-line take|The headline number is the parameter count. The important number is the cluster it ran on, because that one doesn't depend on an export license.

Source: Leiphone (2026-11-02)

Foundation ModelsTencent's Hunyuan ships Hyra 2.0, an AI system that builds research tools

Tencent Hunyuan released Hyra 2.0, reframing its research agent from solving scientific problems directly to having general agents build and upgrade the models, workflows and software used in research. Tencent's phrasing: moving from answering problem sets to manufacturing tools.

What it means for you|Tencent Hunyuan released Hyra 2.0, reframing its research agent from solving scientific problems directly to having general agents build and upgrade the models, workflows and software used in research.

One-line take|Reusable tooling is how a research agent compounds. One-off solutions to one-off problems are a demo budget, not a product.

Source: ITHome (2026-11-02)

Foundation ModelsGPT-6.2 Astra enters grey testing as Claude Mythos 5.5 surfaces the same day

GPT-6.2-Astra began limited testing in early November, with Claude Mythos 5.5 appearing in leaks on the same day. QbitAI reported that weights apparently tied to the release appeared on Hugging Face and were downloaded about 5,000 times before coming down.

What it means for you|GPT-6.2-Astra began limited testing in early November, with Claude Mythos 5.5 appearing in leaks the same day; weights apparently tied to the release were downloaded about 5,000 times before coming down.

One-line take|Two frontier releases inside 24 hours is a schedule, not a coincidence. The 5,000 downloads are the part of the launch nobody planned.

Source: QbitAI (2026-11-02)

Foundation ModelsLi Jiaqi sells 2 trillion tokens in half an hour

On the evening of November 2, Li Jiaqi's livestream sold 100,000 Qwen Office personal monthly cards in 30 minutes, roughly 2 trillion tokens of usage. It was the first Singles' Day where a token package sat in the same slot as lipstick.

What it means for you|On the evening of November 2, Li Jiaqi's livestream sold 100,000 Qwen Office personal monthly cards in 30 minutes, roughly 2 trillion tokens of usage.

One-line take|Tokens have become a consumer good with a discount window. That changes how labs think about demand forecasting, and it puts a live price on inference.

Source: Leiphone (2026-11-02)

Foundation ModelsIDC expects 60% of enterprise AI workloads on open models by 2028

IDC's FutureScape 2027 cycle, covering 250 predictions across 25 research areas, projects that enterprises will run 60% of AI workloads on open-source models by 2028. The forecast assumes cost pressure keeps pushing buyers off closed APIs for the bulk of routine work.

What it means for you|IDC's FutureScape 2027 cycle projects that enterprises will run 60% of AI workloads on open-source models by 2028, assuming cost pressure pushes buyers off closed APIs.

One-line take|Sixty percent of workloads is not sixty percent of value. Open models will take the volume and leave the frontier margins where they are.

Source: ITHome (2026-11-02)

Foundation ModelsChina's model price war is over, and the cost war just started

A TMTPost analysis traces the arc: DeepSeek-V2 pushed input pricing to ¥1 per million tokens in May 2024, ByteDance, Alibaba and Baidu matched within a week, and Baidu made ERNIE Bot free in April 2025 before pricing ERNIE 4.5 at 1% of GPT-4.5. By the second half of 2025, CAICT measured a normalized average of ¥6.8 per million tokens.

What it means for you|A TMTPost analysis traces DeepSeek-V2 pushing input pricing to ¥1 per million tokens in May 2024, rivals matching within a week, and CAICT measuring a normalized average of ¥6.8 per million tokens by late 2025.

One-line take|The price of a token stopped being the story when it hit single-digit yuan. What matters now is the cost of the electricity, the scheduler, and the retries.

Source: TMTPost (2026-11-02)

AI SoftwareByteDance plans to spin off its healthcare unit Xiaohé

ByteDance intends to split off Xiaohé Health, the medical business it founded in 2020, according to Z Finance. Six years in, the scope of the carve-out, any outside funding, and how much control ByteDance keeps are all undisclosed, and the company has not commented.

What it means for you|ByteDance intends to split off Xiaohé Health, the medical business it founded in 2020, according to Z Finance; the scope, outside funding, and retained control are undisclosed.

One-line take|Six years and a spin-off is the polite version of an exit. Healthcare rewards patience ByteDance has never had to spend.

Source: Huxiu (2026-11-02)

AI SoftwareAlibaba Cloud launches Qovest, an AI research desk for institutional investors

Qovest takes a one-sentence research request and runs the whole loop: pulling data, gathering filings and news, analyzing, and producing a report, then re-running it on a schedule. Unlike a chat session, results accumulate in a research space so the next pass starts where the last one ended.

What it means for you|Qovest takes a one-sentence research request and runs the whole loop: pulling data, gathering filings and news, analyzing, producing a report, then re-running it on a schedule.

One-line take|Sell-side research is the most exposed white-collar job in China's AI rollout. The moat was never the model, it was the recurring update cycle, and that is exactly what got automated.

Source: ITHome (2026-11-02)

AI SoftwareTencent Cloud says finance agents have left the trial phase

At Tencent's Global Digital Ecosystem Summit on October 28, cloud VP Hu Liming said half of what he pitched to financial clients a year ago was cloud infrastructure and now almost none of it is. Sales conversations have moved to agents, and the buyers are asking about outcomes rather than model capability.

What it means for you|At Tencent's Global Digital Ecosystem Summit on October 28, cloud VP Hu Liming said half of what he pitched to financial clients a year ago was cloud infrastructure and now almost none of it is.

One-line take|When the sales deck changes before the product does, you are watching a budget line move, not a technology mature.

Source: Leiphone (2026-11-02)

AI SoftwareAnt Digital takes its medical Agentar to Wuzhen's top-ten debut list

Ant Digital Technologies released a healthcare edition of its Agentar platform and was named among the 10 headline debuts at the 2026 Internet Light Expo, held alongside the World Internet Conference in Wuzhen. The category rewards deployed systems over prototypes.

What it means for you|Ant Digital Technologies released a healthcare edition of its Agentar platform and was named among the 10 headline debuts at the 2026 Internet Light Expo in Wuzhen.

One-line take|Hospital deployment is the least forgiving agent market in China: audit trails, liability, and a doctor who will turn it off after one bad answer.

Source: AI Era (2026-11-02)

AI SoftwareAlipay's AHA wins a Wuzhen scenario award for linking agents across devices

Alipay's Agent Hub Access scheme, which lets agents hand tasks between phones, cars and other terminals, took the "Xinyao" scenario award at the 2026 World Internet Conference. The award specifically weighs real deployments and whether they can be copied elsewhere.

What it means for you|Alipay's Agent Hub Access scheme, which lets agents hand tasks between phones, cars and other terminals, took the "Xinyao" scenario award at the 2026 World Internet Conference.

One-line take|Cross-device handoff is the unglamorous plumbing that decides whether an assistant is a feature or an operating layer.

Source: Leiphone (2026-11-02)

📖 The full EastVoice issue on the Chinese forum

EastVoice — five minutes a day, the global AI shift
Physix Frontier (Shenzhen) · covering foundation models, AI software, humanoid robots, autonomous driving and physical AI