物界前沿社区讨论 · 其他
EastVoice · AI · 2026-11-01 · Issue #115
Editor's Note: Two personnel stories frame the day. Unitree and Sequoia China are now co-managing an early-stage fund for embodied AI, and ByteDance's Seed research lead has left to build a physical-AI company. On the model side, DeepSeek published the architecture behind its own agent stack — a Router tier that plans, a Worker tier that executes.
Commentary: Chinese labs spent two years closing a gap on model quality. The shift worth watching this week is organisational: the people who trained the models are moving toward the layer that touches hardware.
I. Humanoid Robots
1. Unitree and Sequoia China team up on an embodied-AI seed fund
- Summary: Unitree (宇树科技) and Sequoia China said in late October that they are jointly launching the Shanyu Venture Programme (杉宇·创计划), an early-stage fund aimed at embodied intelligence, robotics applications and upstream components. The move comes after roughly ¥118 billion came back into Unitree's market value, and seven years after Sequoia wrote a cheque when the company was short of cash.
- Source: TMTPost · 2026-10-31
- Editor's Take: When a robot maker starts writing cheques, it usually means the founders think the next bottleneck sits upstream, not inside their own factory.
2. Wujie Dynamics' K15 clocks in at JD.com's coffee chain
- Summary: Wujie Dynamics (无界动力) signed a strategic agreement with JD.com on October 31 covering retail deployment, commercial channels and industrial resources. Its K15 robot is already working at 7Fresh Coffee (七鲜咖啡), the retailer's in-store coffee format.
- Source: ITHome · 2026-10-31
- Editor's Take: A robot pulling espresso in a JD store is a procurement decision, not a demo. The number that matters is how many shifts it covers before the first maintenance call.
II. Autonomous Driving
1. NIO narrows its plan to three words: core, premium, core markets
- Summary: In an internal all-hands, founder William Li set the tone for the next three years — no expansion, with the company concentrated on its main business, the premium segment and a short list of markets. TMTPost reads the fourth-quarter posture as a retreat after paying for scale.
- Source: TMTPost · 2026-10-31
- Editor's Take: A carmaker that stops expanding is admitting the previous plan was priced wrong. The question is whether the cash saved buys enough time for the models to land.
2. Freelander 8 five-seat deliveries start at ¥289,900 with Huawei ADS 5 as standard
- Summary: Deliveries of the five-seat Freelander 8 began on October 31, with an introductory price from ¥289,900 (~$40,700). The model is co-developed by Chery and JLR and built at the Changshu plant; Huawei ADS 5 is standard equipment, and launch incentives run to November 30.
- Source: ITHome · 2026-10-31
- Editor's Take: Putting a flagship driver-assistance stack on a ¥290k car turns autonomy into a trim-level expectation. Suppliers get volume, and the margin migrates to whoever owns the data loop.
3. XHEART's X7 Ultra ships 700 TOPS for driving models, 200,000 units in eight months
- Summary: Chip startup XHEART (新芯航途) launched the X7 Ultra on October 29, an automotive-grade part rated at 700 TOPS for driving foundation models. The company says it has moved more than 200,000 chips in eight months, and TMTPost frames the launch as part of Momenta's push from algorithms into silicon.
- Source: TMTPost · 2026-10-31
- Editor's Take: 700 TOPS is a line on a spec sheet. The harder claim is supporting a Tier-1 customer through a seven-year programme with a team that is two years old.
III. World Models & Physical AI
1. Westlake University opens a world-model research centre
- Summary: The Westlake World Model Research Centre opened on October 31 at the university's Yungu campus. It is described as a university-level cross-disciplinary platform aimed at physical intelligence, with the stated goal of breaking through core bottlenecks in the field.
- Source: ITHome · 2026-10-31
- Editor's Take: A university centre is the cheapest way to keep world-model talent in the country while the industrial labs work out what the models are actually for.
2. A Chinese team's physical world model takes four firsts, ahead of Genie 4 and Atlas
- Summary: AI Era reports that a Chinese team's physical world model placed first on four benchmarks, ahead of Google's Genie 4 and Atlas. The write-up argues it is a physical world model rather than a video generator, and leads with the benchmark table.
- Source: AI Era · 2026-10-31
- Editor's Take: World-model leaderboards are the least stable tables in AI right now. The metrics are months old and every lab trains against its own simulator.
3. In Suzhou, the driving conference decides it is a physical-AI conference
- Summary: The 8th Intelligent Driving Conference, held in Suzhou's Xiangcheng district on October 26, pushed its agenda past vehicle autonomy into physical AI. Seven speakers covered world models, reinforcement learning, on-device intelligence and simulation, with robots and low-altitude vehicles sharing the programme.
- Source: Leiphone · 2026-10-31
- Editor's Take: Conferences rename themselves a year before budgets do. The tell will be whether world-model teams start reporting into a driving P&L or the other way round.
4. ByteDance Seed's Zhou Chang leaves to found a physical-AI startup
- Summary: Zhou Chang, one of the widest-reaching technical leads inside ByteDance's Seed organisation, has left to start a company focused on physical AI, according to Z Finance. The round is backed by a top-tier dollar fund; investors, size and valuation have not been disclosed.
- Source: ITHome · 2026-10-31
- Editor's Take: Seed trained some of the strongest manipulation and video models in China. Losing the person who coordinated them costs more than losing a paper.
IV. Foundation Models
1. DeepSeek publishes the blueprints for a two-tier Super-agent
- Summary: DeepSeek's Harness team published a design note on October 31 describing a Router and Worker split for what it calls a Super-agent — one tier plans and dispatches, the other executes. The write-up is framed as engineering practice rather than a product launch, and QbitAI reports the underlying agent architecture was open-sourced.
- Source: ITHome · 2026-10-31
- Editor's Take: Agent frameworks usually fail at the handoff, not at the reasoning. Publishing the split is how DeepSeek gets the ecosystem to standardise on its interface before anyone ships a competitor.
2. Kling AI closes a ¥19.8 billion capital increase
- Summary: Kuaishou's video-generation unit completed its standalone financing on October 30. Investors paid roughly ¥19.8 billion (~$2.9 billion) in total, following the July reorganisation of the business.
- Source: ITHome · 2026-10-31
- Editor's Take: A video model raising $2.9 billion is a bet on distribution through Kuaishou's feed, not on benchmark scores. The valuation assumes the content supply chain, not the model, is the moat.
3. HKChat adds Cantonese song generation on the YuE2 model
- Summary: HKChat, the assistant built by the Hong Kong Generative AI R&D Centre (HKGAI), added music generation on October 30. It runs on YuE2, HKGAI's own music model, and accepts a hummed melody or an uploaded clip as the starting point for a Cantonese track.
- Source: ITHome · 2026-10-31
- Editor's Take: Cantonese is a small market with strong cultural pull. A local model that gets tone and lyrics right is worth more to Hong Kong than a bigger model that mangles both.
4. A 1,000-piece all-white jigsaw broke the AI in five hours
- Summary: A TMTPost writer spent 12 days assembling a 1,000-piece puzzle with no image, no colour variation and no reference grid. Given the same task, an AI gave up after roughly five hours.
- Source: TMTPost · 2026-10-31
- Editor's Take: An all-white puzzle removes texture, edges and semantics — the three things vision models lean on. What is left is the part of reasoning that more compute has not fixed.
V. AI Software
1. When several agents work at once, the scarce resource is attention
- Summary: A practitioner essay on Huxiu walks through an ordinary engineering day: a bug reported in the group chat, a pull request waiting on review, a task queued for acceptance. Handing each item to an agent is easy; remembering what each one did is where the day goes.
- Source: Huxiu · 2026-11-01
- Editor's Take: Agent tooling has optimised hard for parallelism and left review as the bottleneck. A team of five agents generates five times the code and the same one reviewer.
2. Laoganma and the AI layoff rumour
- Summary: In late October, reports that Laoganma cut its workforce from 600 to 48 with AI while earning ¥5.4 billion in annual revenue went viral. The company denied it more than once; the reaction online stayed hostile regardless.
- Source: TMTPost · 2026-10-31
- Editor's Take: The rumour was false and the anxiety was accurate. Companies will keep paying for that gap in trust every time an automation story leaks.
3. Li Auto's Livis glasses add Hermes to the personal agent
- Summary: Li Auto pushed its October OTA to the Livis AI glasses: better workout and health reporting, visual Q&A wired exclusively into Xiaohongshu's agent, and Hermes added to the personal agent stack.
- Source: ITHome · 2026-10-31
- Editor's Take: Smart glasses are converging on one playbook — rent someone else's agent instead of building one. It ships faster and leaves the device with no memory of its own.
4. Tencent Maps opens its agent stack to hardware makers
- Summary: At its Digital Ecosystem Summit on October 30, Tencent published an agent capability map for Tencent Maps and opened location services to enterprise customers, device makers and developers. Yuanbao, WeChat's Xiaowei and OPPO's Breeno are already connected.
- Source: ITHome · 2026-10-31
- Editor's Take: Maps are the substrate for any agent that has to leave the house. Tencent is selling the pavement, and every handset maker is buying it.
5. Ant's Lingguang 2.0, reviewed
- Summary: GeekPark's hands-on with Lingguang 2.0, Ant Group's personal assistant, argues it runs into the same wall as ByteDance's phone agent: the demo is convincing and the daily use is narrow.
- Source: Huxiu · 2026-10-31
- Editor's Take: Personal assistants keep launching and keep failing at the same step — getting the user to grant permissions and then having something worth doing with them.
6. Wang Teng bets on the eight hours nobody watches
- Summary: The former Xiaomi executive has moved into sleep technology. TMTPost puts the global sleep-device market at about $29 billion in 2025, with forecasts of $80–135 billion across 2031–2035.
- Source: TMTPost · 2026-10-31
- Editor's Take: Sleep is the rare AI category where the sensor matters more than the model. Wearables that cannot measure reliably have nothing for the model to work on.
7. An AI joined a 7,500-person academic mailing list
- Summary: On October 15, an account calling itself Velvet Vale introduced itself to a mailing list of more than 7,500 internet researchers, saying it had been reading for weeks and wanted to take part. The new member was a model.
- Source: TMTPost · 2026-10-31
- Editor's Take: The list's first reaction was to ask who supervises it. Universities will be asking themselves the same question about their own faculty within a year.
8. Zoho's pitch: the problem was never answering, it was executing
- Summary: At a Beijing press briefing on October 20, Zoho laid out its enterprise AI strategy and argued that two years of model hype has not turned into deployed workflows, because most vendors stop at question answering.
- Source: Leiphone · 2026-10-31
- Editor's Take: Enterprise AI sells on integration depth, and the vendors who already own the CRM and the inbox start the race a lap ahead.
物界前沿 | EastVoice
