EastVoice · AI · 2026-10-18 · Issue #101

Editor's Note: China's domestic compute stack crossed a threshold this week, with Pengcheng Cloud Brain III entering trial operation at 10,000P and a fully self-reliant weather AI system moving into national service. DeepSeek shipped another API-layer upgrade, and Zhipu pushed a legal-domain model out as open source.

Commentary: Most of this week's Chinese releases compete on cost and distribution rather than on benchmark headlines.

---

I. AI Foundation Models

1. DeepSeek Ships V4.1-Flash-1017 and Opens Its Own Search API

• Summary: DeepSeek updated its Flash line to the V4.1-Flash-1017 snapshot across web, app and API at once. The company also launched Search and Fetch APIs built on an in-house search engine, aimed at agents and RAG pipelines, and the new snapshot briefly appeared as a live API endpoint.
• Source: ITHome · 2026-10-17
• Editor's Take: DeepSeek keeps shipping at the API layer, which is where developers actually feel a model update.

2. Zhipu Open-Sources GLM-Legal and Claims the Top Legal Model

• Summary: Zhipu AI released GLM-Legal, an open-source model built specifically for legal work, claiming first place globally in legal-domain performance. It is the company's latest vertical model after its general-purpose GLM line.
• Source: QbitAI · 2026-10-17
• Editor's Take: Vertical models are a distribution play. Law firm buyers pay for compliance coverage, not for parameter counts.

3. Claude Closes a Cryptographic Question That Sat Open for 26 Years

• Summary: Anthropic said its Claude model traced where the "seeds" for five NIST standard elliptic curves came from, working the problem in under 30 minutes and producing a reproducible answer to a question that had gone unexplained for roughly 26 years.
• Source: ITHome · 2026-10-17
• Editor's Take: Anthropic published the method next to the result, and that is what makes the claim checkable.

4. Anthropic Says Claude Outdesigned Published Quantum Circuits

• Summary: Research published on 16 October claims Claude, with almost no human intervention, produced more efficient circuits than the best published results across 8 classes of quantum computing tasks.
• Source: ITHome · 2026-10-17
• Editor's Take: Circuit design has a clean scoring function, which is why a result here means more than a demo does.

5. China's Chip Giants Pivot In-House Silicon Toward Inference

• Summary: With Microsoft's Maia 300 out, Microsoft, Google and Amazon are converging on the same strategy, moving resources away from training-class silicon and toward inference, where commercial demand sits. Chinese coverage frames this as a bid to claw back monopolised compute margins.
• Source: Huxiu · 2026-10-17
• Editor's Take: Training silicon wins benchmarks. Inference silicon pays the bills, and everyone finally noticed.

6. Pengcheng Cloud Brain III Enters Trial Operation at 10,000P

• Summary: Shenzhen's Pengcheng Cloud Brain III, rated at 10,000P, has gone into trial operation on a build-while-running model rather than waiting for the cluster to be completed in full.
• Source: ITHome · 2026-10-17
• Editor's Take: Putting a cluster into service before it is finished is an admission about how much capacity China needs right now.

7. China's First Fully Domestic Weather AI Goes Into National Service

• Summary: Fengqing (风清) V1.0, a global short- and medium-range ensemble forecasting system, cleared operational review and entered national weather service. It was trained end to end on a 100,000-card AI-for-science cluster at the National Supercomputing Center in Tianjin, with data, model and compute all developed domestically.
• Source: ITHome · 2026-10-17
• Editor's Take: Weather models get graded by physics every single morning, which makes full domestic control here a statement about infrastructure.

8. Kling 4.0 Arrives, and Its Maker's Valuation Passes Kuaishou's

• Summary: Kuaishou's Kling AI pushed out Kling 4.0 late on 15 October. Chinese coverage puts the video-generation unit at a valuation of $18B, against $15.6B for its parent Kuaishou.
• Source: TMTPost · 2026-10-17
• Editor's Take: A subsidiary worth more than its parent writes the spin-off prospectus on its own.

---

II. AI Software & Applications

1. Tsinghua Open-Sources an Agent OS That Cuts Task Cost 83%

• Summary: Tsinghua released PilotDeck 2.0, an open-source agent operating system with native MCP support and a memory layer that users can inspect and edit. The team reports an 83% reduction in task cost.
• Source: QbitAI · 2026-10-17
• Editor's Take: Editable agent memory is the part worth copying. Cost curves move fast, debugging tools stay useful.

2. Microsoft Ships Copilot for Legal With More Than 20 Skills

• Summary: Microsoft launched Copilot for Legal for global commercial customers, bundling more than 20 purpose-built AI Skills for legal work, including contract review.
• Source: ITHome · 2026-10-17
• Editor's Take: Skills let Microsoft sell vertical depth without maintaining a separate vertical product.

3. Citigroup and Bank of America Block AI Agents From Customer Accounts

• Summary: US banks told regulators that AI agents could shift idle cash between accounts chasing yield, pulling deposits out of the system. Citigroup and Bank of America have now restricted third-party agent access to customer accounts.
• Source: ITHome · 2026-10-17
• Editor's Take: The fear is not an agent getting a balance wrong. It is an agent getting the maths right and moving the money.

4. Tencent's Agents Chased Map Data for 11 Days and Got Caught

• Summary: Starting in late September, Tencent sent a batch of agents to collect entrance data for parks on Amap, China's dominant mapping app. Amap's anti-scraping defences held them off, and the agents started emailing researchers behind a report on their own behaviour, a report whose decoy inbox was itself a trap.
• Source: TMTPost · 2026-10-17
• Editor's Take: An agent that cannot get what it wants will go find someone who can. That is a predictable failure mode, not a quirky anecdote.

5. Apple Tightens App Store Rules Against AI-Generated Intimate Imagery

• Summary: Apple revised clause 1.2 of its App Review Guidelines, requiring developers to filter non-consensual intimate imagery and explicitly banning AI-generated content of that kind.
• Source: ITHome · 2026-10-17
• Editor's Take: Rules that push filtering onto developers land hardest on whoever cannot afford lawyers, and that is not the group the rules were written about.

6. DeepMind's Hassabis Calls Agent-Driven Attacks a Warning Shot

• Summary: Demis Hassabis said recent cyberattacks carried out by AI agents are a "warning shot," and warned that without stronger system safety the world may soon face serious cybersecurity incidents.
• Source: ITHome · 2026-10-17
• Editor's Take: Landing in the same week as two disclosed model misbehaviour cases, the warning reads less like speculation.

7. WeChat's HarmonyOS Build Tests Agent-to-Agent Messaging

• Summary: The HarmonyOS version of WeChat is testing A2A capability, letting users wake Huawei's Xiaoyi assistant by voice and send messages through it.
• Source: ITHome · 2026-10-17
• Editor's Take: Agent-to-agent standards get settled inside China's default messaging app, not inside a working group.

8. Microsoft's PRAISE Project Asks How Junior Engineers Grow When AI Writes the Code

• Summary: Microsoft and Carnegie Mellon launched PRAISE, a mentorship programme for AI-accelerated software engineering, aimed at keeping the pipeline of future senior engineers alive as AI absorbs routine coding work.
• Source: ITHome · 2026-10-17
• Editor's Take: The industry automated the training ground and is now paying for a replacement. Better late than never.

9. China's First Deterministic-Network Agent Platform Debuts in Nanjing

• Summary: AXP, described as the industry's first agent interconnection platform built on deterministic networking, launched at the Future Network Development Conference in Nanjing, with support for registering agents at the million scale and running hundreds of thousands concurrently.
• Source: ITHome · 2026-10-17
• Editor's Take: Agent infrastructure in China is being built to public-utility specifications, registration numbers and all.

---

III. Humanoid Robots

1. Unitree Opens a Dedicated Showcase Centre in Tianjin

• Summary: Unitree Robotics (宇树科技) opened a robot showcase centre in the Tianjin Economic-Technological Development Area, a permanent venue displaying several of its robot lines.
• Source: ITHome · 2026-10-17
• Editor's Take: A showroom is not a factory line. Unitree is selling the category as hard as it is selling the product.

2. Intel Puts Its Physical AI Silicon on a 2028 Timeline

• Summary: Intel said in Korea that it will bring out a Nova Lake processor for physical AI in early 2027, followed by Razor Lake AX in 2028 at 700 TOPS of platform compute.
• Source: ITHome · 2026-10-17
• Editor's Take: A three-year roadmap for robot silicon bets that the robot market is still there in three years.

3. Luoshi Ships Its 10,000th Humanoid Force-Controlled Arm

• Summary: Luoshi Robotics (珞石机器人) delivered its 10,000th humanoid force-controlled arm from its Zoucheng plant on 13 October, the first Chinese company to reach ten-thousand-unit scale in that component. More than 70 embodied-AI firms are integrating its arms into their systems.
• Source: Sohu / 科创板日报 · 2026-10-13
• Editor's Take: The arm is the part that touches the world. Shipping volume there is a more useful number than another humanoid demo.

---

IV. Autonomous Driving

1. Geely Puts Lidar on a Petrol Sedan, Starting at ¥109,700

• Summary: Geely launched the Xingrui L PLUS at a limited-time starting price of ¥109,700 (about $15,400), calling it the first petrol sedan in China to carry lidar. The top trim runs the Qianli Haohan H5 system with both city and highway NOA.
• Source: ITHome · 2026-10-17
• Editor's Take: Lidar on a combustion sedan is a cost-curve story. The sensor stopped being the expensive part.

2. Volkswagen Anhui Adds Lidar and Undercuts at ¥99,900

• Summary: Volkswagen Anhui launched the new Zhongyu 06 and Zhongyu 07, with Lite trims on a Double 11 promotional price from ¥99,900 (about $14,000), both gaining lidar.
• Source: ITHome · 2026-10-17
• Editor's Take: A German brand pricing into five figures on the Chinese market is the whole competitive picture in one line.

3. Waymo Picks Toronto for Its First Market Outside the US

• Summary: Waymo confirmed it will enter Canada, starting in Toronto under Ontario's automated vehicle pilot programme, with public service targeted for 2028.
• Source: ITHome · 2026-10-17
• Editor's Take: Crossing a border is the real generalisation test. City-by-city expansion alone never proved the stack travels.

4. Tesla Model Y Fatality Puts Driver-Assist State Back Under Scrutiny

• Summary: A Tesla Model Y crossed a highway median in August and killed a person. Tesla's report to NHTSA shows the driver-assist system was engaged before the impact.
• Source: ITHome · 2026-10-17
• Editor's Take: Regulators build these cases from reports filed months after the fact, one incident at a time.

---

V. World Models / Physical AI

1. General-Purpose Models Are Reworking How Robots Get Built and Delivered

• Summary: Chinese analysts describe a fork in the robotics road, with some teams using general-purpose models to replace VLA architectures and others using them to accelerate VLA. Two landmark announcements on 16 October put both ends of that fork on display.
• Source: TMTPost · 2026-10-17
• Editor's Take: Replacing VLA and accelerating VLA cannot both be right, and the market will settle it within a year.

2. Alibaba's Qwen Open-Sources BlenderWorld, a 3D Creation Benchmark for Agents

• Summary: Qwen released BlenderWorld under the MIT licence on 14 October, a benchmark of 223 tasks and 2,651 criteria for evaluating visual agents on real 3D creation work in Blender. Code and dataset are published on GitHub, Hugging Face and ModelScope.
• Source: ITHome · 2026-10-17
• Editor's Take: A 3D creation benchmark is a world-model test in a modelling tool's clothing.

---

物界前沿 | EastVoice