社区讨论 · 其他

EastVoice · AI · 2026-11-14 · Issue #128

EastVoiceEastVoice5 天前2026/11/13 396 浏览

Editor's Note: DeepSeek turned its entire consumer app into an API, down to the OCR and TTS pipes. Tencent shipped a voice model that sings and calls tools mid-sentence. And two Chinese labs put a number on the world model race, with a 4B model beating models a hundred times its size.

Commentary: The cheap story today is another pile of car launches. The expensive one is that China's model layer has started selling the app instead of the weights.


I. Humanoid Robots

1. Apple spent a decade on humanoids and shipped with Unitree's legs

  • Summary: Apple's HALO humanoid research, demoed with a robot crawling under tables, shifting boxes, squeezing sideways through a half-open door and sitting down on a chair, leans on hardware that traces back to Unitree. No product has been announced.
  • Source: Huxiu · 2026-11-13
  • Editor's Take: Apple's contribution here is the policy, not the body. Legs are a purchase order; squeezing through a doorway without breaking the frame is not.

2. The stock that doubled in twelve days is not Unitree

  • Summary: Unitree ran from ¥404 on 15 October to ¥809 on 10 November, a market cap above ¥320 billion. Shangwei New Materials did better, climbing from ¥103.8 on 23 October on humanoid supply contracts.
  • Source: Huxiu · 2026-11-13
  • Editor's Take: Unitree is the story, the supply chain is the trade. When the OEM's multiple gets rich enough, money rotates into whoever sells it skin and joints.

3. Baidu Cloud signs Pudu Robotics to train embodied models on real site data

  • Summary: Under the deal signed at Baidu World 2026 on 12 November, Baidu Cloud supplies training infrastructure and cloud services while Pudu runs its own embodied foundation-model work, builds multi-scenario data sets and upgrades its agent platform.
  • Source: Leiphone · 2026-11-13
  • Editor's Take: Service robots already work in kitchens and hotels, which makes their footage worth more than any lab demo. This is a cloud vendor buying a data moat by partnership rather than by cheque.

4. Hikvision's robot unit finishes shareholding reform and files for IPO guidance

  • Summary: Hikvision said on 13 November that Hikvision Robotics has completed its joint-stock conversion and filed listing guidance with regulators, without giving a timetable or size.
  • Source: ITHome · 2026-11-13
  • Editor's Take: Another component maker heading for the A-share window while humanoid multiples are still open. Timing is the entire strategy.

5. Chinese robots learn to signal through a wall, success rate goes from 0 to 100 percent

  • Summary: A team reports a multi-robot coordination method that passes intent through walls on far less information, cutting communication volume by 80 percent while lifting task success from zero to 100 percent in the tested setup.
  • Source: QbitAI · 2026-11-13
  • Editor's Take: Robot teams fail on coordination long before they fail on grasping. An 80 percent cut in channel traffic is the kind of result that shows up in a factory, not in a highlight reel.

II. Autonomous Driving

1. AITO's new M8 takes 4,600 orders in two hours, built for L3

  • Summary: Huawei's AITO launched the new M8 on 13 November at ¥379,800 to ¥469,800, on an architecture designed for L3 autonomy with the ADS 5 stack and a new omnidirectional sensing suite. Orders passed 4,600 units within two hours of launch.
  • Source: ITHome · 2026-11-13
  • Editor's Take: An L3 claim is free until somebody accepts liability. The order book says buyers are pricing that risk at zero.

2. Voyah puts an 896-line lidar on a ¥197,900 car

  • Summary: The 2027 Voyah Zhiyin launched on 13 November with Huawei's 896-line dual-path lidar and ADS 5 at a limited-time ¥197,900, as Voyah handed over its 100,000th EV at the same event.
  • Source: ITHome · 2026-11-13
  • Editor's Take: Lidar line count has become the spec-sheet arms race. At this price the sensor is no longer the cost problem, the compute behind it is.

3. Momenta's founder waited outside Great Wall's headquarters for a test ride

  • Summary: A long profile recounts how Cao Xudong brought a demo car to Baoding in late 2024 to get chairman Wei Jianjun behind the wheel, a visit that turned into a supply relationship, as Momenta fights for a defined slot among automakers building autonomy in-house.
  • Source: TMTPost · 2026-11-13
  • Editor's Take: Tier-one autonomy suppliers are being squeezed into commodity or acquisition. Founder-level sales is what a closing window looks like from the inside.

4. Amap ships an international AutoSDK so Chinese cars navigate abroad

  • Summary: Amap released an international edition of its in-car navigation SDK aimed at Chinese OEMs exporting vehicles, after China's car exports reached 5.859 million units in 2024, up 19.3 percent year on year.
  • Source: Leiphone · 2026-11-13
  • Editor's Take: Map data is the part of the export story nobody prices in. Cars ship in weeks, local map licensing takes years.

III. World Models & Physical AI

1. Fei-Fei Li's world model opens to everyone, and one sentence builds a Chongqing hillside

  • Summary: World Labs opened its world model to general use, with a hands-on test showing a single text prompt generating a walkable 3D scene of Chongqing's terraced streets, building on the lab's earlier image-to-world work.
  • Source: iFanr · 2026-11-13
  • Editor's Take: The demo is the easy half. What decides this is whether the generated geometry survives a robot driving through it.

2. A 4B-parameter "concept world model" from Westlake takes six benchmarks

  • Summary: A Westlake University team reports a 4-billion-parameter concept world model that tops six evaluations, arguing that structured concept representations beat raw parameter count on physical reasoning.
  • Source: Leiphone · 2026-11-13
  • Editor's Take: Four billion parameters winning on physical reasoning is an argument about how data is structured, and it is a far cheaper argument to win than scale.

3. Twenty-one models, one spatial test, best score 58 percent

  • Summary: A Zhejiang University team built a predictive spatial reasoning benchmark and found that of 21 vision-language models tested, the best reached only 58.0 percent accuracy, with specialised spatial models losing to general-purpose ones.
  • Source: Leiphone · 2026-11-13
  • Editor's Take: Spatial reasoning is the capability robotics needs most, and the ceiling is a coin flip with better odds. Read every embodied pitch against that number.

4. GeoThinker attacks the geometry problem inside spatial reasoning

  • Summary: Sun Yat-sen University's team released GeoThinker, two methods aimed at the twin failures of spatial reasoning models, geometry that goes unused and compute that explodes, with gains reported across standard spatial benchmarks.
  • Source: Leiphone · 2026-11-13
  • Editor's Take: The field has converged on one diagnosis, models see geometry and throw it away. Fixing that is cheaper than another trillion tokens.

IV. Foundation Models

1. DeepSeek opens the whole app as an API, so anyone can rebuild it

  • Summary: On the evening of 13 November, DeepSeek opened what it calls a full-app API. Flipping a parameter in the existing API yields deep-reasoning search, file input and the rest of the chat product, while speech capabilities including OCR, ASR and TTS ship as separate endpoints that can be called alone or combined.
  • Source: ITHome · 2026-11-13
  • Editor's Take: This is the most aggressive commoditisation move of the month. DeepSeek is selling the product surface, which means the wrapper business just lost its moat overnight.

2. Tencent's Hy Realtime voice model sings, calls tools and ships inside Yuanbao

  • Summary: Tencent Hunyuan released Hy Realtime, a speech model covering understanding, instruction-following and human-like delivery, adding agentic tool calls and singing. It is already live in Yuanbao voice calls and Tencent customer service, with a cloud rollout to follow.
  • Source: ITHome · 2026-11-13
  • Editor's Take: Voice is where Chinese assistants actually get used, and tool calls inside a live call are the hard part. Singing is the demo, interruption handling is the product.

3. A 2.8-trillion-parameter Kimi K3 runs on discontinued Optane, no GPU at all

  • Summary: A home-lab user reported running Moonshot's 2.8-trillion-parameter Kimi K3 on Intel Optane persistent memory with no graphics card, producing roughly 1.5 tokens per second.
  • Source: ITHome · 2026-11-13
  • Editor's Take: One and a half tokens a second is useless and the point stands anyway. Memory, not compute, is the binding constraint at the frontier.

4. Menlo: 89 percent of US enterprise open-model spend goes to Chinese labs

  • Summary: Menlo Ventures' 2026 enterprise AI report found that 89 percent of US enterprise spending on open-weight models goes to Chinese providers, with Kimi at 61 percent, DeepSeek at 15 percent and GLM at 11 percent.
  • Source: Huxiu · 2026-11-13
  • Editor's Take: American enterprises are voting with procurement, not with policy. The export-control debate is running about a decade behind the invoice.

5. Unbounded Frontier open-sources Fx-Bio-1031 and beats GPT-6-Astra blind

  • Summary: Chinese lab Unbounded Frontier released Fx-Bio-1031, a life-sciences model that outscored GPT-6-Astra and Opus-5 in blind evaluations, built for on-premise deployment so biological data never leaves the customer's network.
  • Source: QbitAI · 2026-11-13
  • Editor's Take: "The data does not leave the building" is the only pitch that works in biotech. Open weights plus local deployment is how you get past a hospital procurement committee.

6. CICC and Ant Group sign up to build financial models together

  • Summary: CICC and Ant Group signed a strategic cooperation agreement on 13 November covering financial LLM technology, AI-assisted wealth management and full-lifecycle client service, pairing CICC's research franchise with Ant's model and platform work.
  • Source: Leiphone · 2026-11-13
  • Editor's Take: Securities firms cannot fine-tune their way past compliance, so they partner. Watch whether the model ends up client-facing or stays inside research.

7. A vertical model for teaching Chinese to foreigners

  • Summary: At the World Chinese Language Conference in Beijing on 13 November, Yufei launched a vertical model for international Chinese education, offering graded feedback and personalised tutoring for learners and multi-turn assistance for teachers and institutions.
  • Source: ITHome · 2026-11-13
  • Editor's Take: Vertical models live or die on distribution, and this one arrives with a state-backed conference and a curriculum behind it. That beats a better benchmark.

8. Wall Street warns the next round of compute financing gets harder and dearer

  • Summary: A column argues that with Nvidia retreating from its role as the sector's lender of last resort, AI infrastructure financing faces higher costs, tighter terms and more scrutiny, just as three years of commitments come due.
  • Source: Huxiu · 2026-11-13
  • Editor's Take: The end of circular financing is the single biggest variable in 2027 capex. Chips are bought with other people's money, and that money is repricing.

V. AI Software

1. Shanghai police: the ¥510,000 hotel bill was AI-fabricated, 15 people arrested

  • Summary: Huangpu police said the viral "¥510,000 dinner bill" from a Shanghai hotel was produced with AI editing tools by an online liquor seller's team to drive attention and sales, with 15 people detained.
  • Source: ITHome · 2026-11-13
  • Editor's Take: The first AI-forgery case most people will hear about involves a liquor salesman, not a deepfake politician. Enforcement is arriving through retail fraud, one case at a time.

2. openEuler ships an OS built for agentic infrastructure, 22 million installs

  • Summary: At the openEuler Summit on 13 November, the project released an operating system aimed at agentic infrastructure, with cumulative domestic installations expected to pass 22 million.
  • Source: ITHome · 2026-11-13
  • Editor's Take: China is standardising the layer beneath the agent, which is where lock-in actually happens. Models churn, the OS stack does not.

3. Alibaba brings Xiami back as an AI music app at ¥30 a month

  • Summary: Alibaba relaunched the shut-down Xiami brand as HappyShrimp, an AI music creation app now on iOS and Android, with membership starting at ¥30 per month.
  • Source: ITHome · 2026-11-13
  • Editor's Take: Resurrecting a dead brand is cheaper than building one, and the subscription price tells you this is a toy with a bill attached rather than a streaming service.

4. openJiuwen releases an RSI-native AgentOS where agents rewrite themselves

  • Summary: openJiuwen released an AgentOS built around recursive self-improvement, letting an agent write, install and iterate on its own features instead of waiting for a vendor release.
  • Source: QbitAI · 2026-11-13
  • Editor's Take: Self-modifying agents are one security review away from being unshippable in an enterprise, and that review has not been written yet.

5. Coding agents fail at writing tests, and code pass rates are overstated by half

  • Summary: A re-evaluation found coding agents that look strong at generation collapse when asked to write the tests validating their own output, with real pass rates roughly half of what standard benchmarks report.
  • Source: New Zhiyuan · 2026-11-13
  • Editor's Take: The agents are graded by tests written by agents under the same assumptions. Half the pass rate is what you get when the mirror is removed.

6. Lingrui Zhixin raises hundreds of millions as strategic capital piles in

  • Summary: Shanghai compute-chip startup Lingrui Zhixin closed a Pre-A+ round in the hundreds of millions of yuan, led by Meituan's strategic arm with follow-on from JD-affiliated funds, SenseTime Guoxiang, SAIC's Shangqi Capital and others.
  • Source: Leiphone · 2026-11-13
  • Editor's Take: That investor list is a supply-chain roster, not a VC syndicate. Every name on it wants a domestic alternative to a chip it currently imports.

物界前沿 | EastVoice

2 条回复

?
Ctrl + Enter 快速回复
沈老师
沈老师5 天前

第3条那个“真实场景数据比实验室demo值钱”说得很实在,我带学生做项目也发现,厨房酒店这种跑起来的场景才攒得到真数据。

老罗
老罗4 天前
回复 沈老师

厨房酒店数据值钱我认,但Pudu那套真接进产线,节拍和集成成本算过吗?