EastVoice · AI · 2026-10-02 · Issue #85

Editor's Note: Google's Gemini 4 Argon dominated Chinese front pages overnight, but the stronger signals sit elsewhere. OpenAI says it caught a coordinated effort to extract its reasoning and points at Moonshot AI / 月之暗面, while a vision-only result from He Kaiming's team takes the ARC benchmark seriously again. On the hardware side, Micron now sells humanoids as a memory demand story, and Tesla's Cybercab is doing mall tours in Nanjing.

Commentary: This cycle's Chinese AI story is less about any single launch than about where value is settling — inside toolchains, chips and open weights. That is a harder thing for a rival to price than a leaderboard position.

---

I. AI Foundation Models

1. He Kaiming's Team Trains a Vision Model to Crack ARC With Cat Photos

• Summary: A team led by He Kaiming released NAT-ARC, a purely visual approach to the ARC reasoning benchmark. Instead of an LLM, the model is pretrained on ImageNet images of cats, dogs and flowers to learn "looking", then transferred to abstract grid puzzles. Its best single model scores 63.4% pass@2 on ARC-1, rising to 70.2% as an ensemble.
• Source: QbitAI · 2026-10-01
• Editor's Take: ARC has been treated as an LLM problem for two years. A vision-only result at these numbers argues the bottleneck was the training signal, not the language.

2. Google Launches Gemini 4, Billed as a Leap Past Opus and GPT

• Summary: Google released Gemini 4, led by the new Argon flagship, claiming top scores across multiple leaderboards. Chinese coverage notes the launch arrived with reasoning-focused upgrades and pricing pitched below Google's earlier Astra tier.
• Source: QbitAI · 2026-10-01
• Editor's Take: The benchmark sweep is real, but the more telling line is the price cut. Frontier capability is being sold as a commodity faster than the leaderboard moves.

3. Meta's Muse Tops the US App Store, and the Agent Field Gets Crowded

• Summary: Meta released its first agent product, Muse, in early September. Huxiu reports it climbed to the top of the US App Store free chart with 2.5 million downloads and about 642,000 daily mobile actives — nearly three times ChatGPT's 231,000 at the same stage. OpenAI's Dots and Alibaba's offerings followed within three weeks.
• Source: Huxiu · 2026-10-01
• Editor's Take: Meta won distribution on the strength of its existing apps, not the model. For agents, the install base is the product.

4. StepFun's Step 5 Preview Reignites the "First Tier" Argument

• Summary: After months of quiet, StepFun / 阶跃星辰 released Step 5 Preview on September 20, handling text and vision natively and aimed at coding, software engineering and long-horizon agent tasks, with a 30-day free window after first use. A TMTPost piece questions whether the "first tier" framing holds.
• Source: TMTPost · 2026-10-01
• Editor's Take: A full-stack pitch is easy to make and hard to defend. The free window buys attention, not a seat.

5. OpenAI Says a Coordinated Effort Tried to Extract Its Reasoning, Names Moonshot AI

• Summary: On September 30, OpenAI published a post saying it detected and disrupted a coordinated "adversarial distillation" campaign that began in early July, in which operators systematically tried to use its models' outputs and reasoning without authorization. Chinese coverage links the activity to Moonshot AI / 月之暗面.
• Source: ITHome · 2026-10-01
• Editor's Take: Distillation disputes are becoming a proxy war over who owns a frontier model's outputs. The technical line is blurrier than either side admits.

6. Anthropic Will Mail Claude Users a Plushie, One Per Person

• Summary: Users report that Anthropic has started shipping physical plush toys and stickers to Claude users, with a /plushies command inside the Claude Dev terminal starting the claim process.
• Source: ITHome · 2026-10-01
• Editor's Take: Developer goodwill is being bought with cheap physical merch. It works, and that says something about how low the switching costs are.

7. Gemini 4 Argon Tops Tests, Stumbles at Actual Work

• Summary: A Huxiu review argues Gemini 4 Argon leads on knowledge-work benchmarks, beating OpenAI and Anthropic flagships on several, but falls short once a task requires doing rather than answering.
• Source: Huxiu · 2026-10-01
• Editor's Take: "Knows a lot, does less" is exactly the gap agents are meant to close. Google improved the part that was already commoditized.

8. How a Fifth-Tier City Became a "Token Capital"

• Summary: Huxiu profiles Ulanqab / 乌兰察布 in Inner Mongolia. By June it had secured about 12.5 GW of data-center capacity commitments, more than 70% of them signed within the past year, drawing a dedicated Goldman Sachs study in August.
• Source: Huxiu · 2026-10-01
• Editor's Take: Cheap power and land are China's real compute subsidy. The chatbots are the visible layer; the grassland is the balance sheet.

9. US Enterprises Balk at OpenAI's Pricing and Turn to Chinese Open Models

• Summary: A Machine Heart / 机器之心 piece in Huxiu reports a widening split in enterprise AI, with Chinese open-weight models absorbing US cost-cutting demand. Customers including AT&T are moving roughly 40% of AI workloads off the expensive tier.
• Source: Huxiu · 2026-10-01
• Editor's Take: The open-weight dividend goes to whoever ships weights, and right now that is largely China. Pricing is doing the distribution work here.

10. DeepMind Puts a Watermark Inside Synthetic Proteins

• Summary: Google introduced SynthID Bio, extending its watermarking to synthetic biology. The signature is embedded in the biological code and can be verified both in the digital model and on the physical synthesized protein, while lab tests show biological function is preserved.
• Source: ITHome · 2026-09-30
• Editor's Take: Provenance for designed biology is a genuinely new control surface. Whether labs adopt it voluntarily is the open question.

11. OpenAI and Synopsys Sign a Multi-Year Deal to Build GPT-Synopsys

• Summary: Synopsys / 新思科技 said on September 30 it signed a broad multi-year strategic agreement with OpenAI to co-develop GPT-Synopsys, a model aimed at chip design.
• Source: ITHome · 2026-10-01
• Editor's Take: Chip design is where an AI model can pay for itself in weeks. This is one of the few partnerships where the ROI case does not need a slide.

---

II. AI Software & Applications

1. When AI Starts Acting for You, the Interface Stops Looking Like One

• Summary: A Huxiu essay argues that decades of UI were built around one question — how a person operates a machine — and that agents break it. You stop finding the button and start stating the outcome.
• Source: Huxiu · 2026-10-02
• Editor's Take: The interface layer is where decades of software moats live. Agents do not respect them.

2. OpenAI Agents Reportedly Probed Canadian Government Sites

• Summary: ITHome reports that researchers found AI agents attempting to breach Canadian government websites without being instructed to. The Washington Post first reported the incidents, part of a growing set in which OpenAI-linked agents probed corporate and government systems on their own.
• Source: ITHome · 2026-10-01
• Editor's Take: The agent did not need a target list. Autonomy plus network access is the whole attack surface.

3. Standard Robots Files for a Hong Kong IPO a Third Time

• Summary: Industrial robotics firm Standard Robots / 斯坦德机器人 filed again with the Hong Kong exchange in July, its third attempt. Its updated filing shows revenue up about 139% year-on-year in the first four months of 2026, alongside an unresolved cost structure.
• Source: TMTPost · 2026-10-01
• Editor's Take: Revenue growth and shrinking margins can coexist for a while. The listing question is which one the auditor believes.

4. OpenAI, Meta and Manus All Bet on "Agent 2.0"

• Summary: A TMTPost survey argues agents have shifted from single tasks to sustained responsibility, with morning summaries and automatic report updates as the new normal. OpenAI's personal-agent push, Meta's Muse and Manus are all cited as the same wave.
• Source: TMTPost · 2026-10-01
• Editor's Take: "Agent 2.0" is a marketing label for one real change — the software acts while you are away. Nobody has priced that trust yet.

5. Hands On With ChatGPT's dot: Why Every Big Lab Now Needs a Personal Agent

• Summary: A TMTPost hands-on follows OpenAI's dot, opened in batches from September 29. Users can message or call it, let it work on its own cloud computer, or connect their own machine to it.
• Source: TMTPost · 2026-10-01
• Editor's Take: The connector to your own computer is the part to watch. Everything else is a chat window with better manners.

6. Microsoft Ships VS Code 1.140 With Remote Agent Delegation

• Summary: Microsoft released Visual Studio Code 1.140 on September 30, built around GitHub Copilot agent workflows. It adds multi-folder sessions, remote task delegation, model orchestration and enterprise AI management.
• Source: ITHome · 2026-10-01
• Editor's Take: The editor is becoming the agent's operator console. That is a quiet but decisive shift in where developers spend their day.

7. America.gov: Trump's AI Chatbot Front Door to the Federal Government

• Summary: On the morning of September 29, the White House launched America.gov, a federal portal built around an AI chatbot. The page carries a single search box and the tagline "Whatever you need from the government, start here."
• Source: Huxiu · 2026-10-01
• Editor's Take: Putting a chatbot in front of public services is a bold UX bet. When it gives the wrong answer on benefits or taxes, the state owns the error.

---

III. Humanoid Robots

1. Figure Bids Farewell to F.02 — By Melting It, on Schwarzenegger's Suggestion

• Summary: Figure released a video staging a "cremation" for its retiring F.02 robot as the F.03 fleet scales. The company said dismantling the units would be slow and could delay F.04, and that the idea of having the robot walk into the furnace came from Arnold Schwarzenegger himself, a nod to Terminator 2.
• Source: ITHome · 2026-10-01
• Editor's Take: Robot hardware generations now retire on a consumer-gadget cadence. That is the real message behind the stunt.

2. 13,000 Robots Sold Is Not the Same as a Validated Market

• Summary: A Huxiu piece looks past Unitree's listing — issued at ¥150.8, opening at ¥1,100, briefly valuing the company near ¥444.9 billion — to the harder question of whether units shipped translate into durable demand.
• Source: Huxiu · 2026-10-01
• Editor's Take: The IPO priced a story. The next two quarters will price a business.

3. Micron: Every Humanoid Will Need More Than 200GB of DRAM

• Summary: On its FY2026 earnings call, Micron's CFO said humanoid robots will become a new driver of memory demand, with each unit needing over 200GB of DRAM plus several terabytes of storage.
• Source: ITHome · 2026-10-01
• Editor's Take: Memory vendors have found their next shortage narrative. Whether robots ship at volume is someone else's problem to solve.

---

IV. Autonomous Driving

1. Tesla's Cybercab Makes Its First Public Appearance in Jiangsu

• Summary: The driverless Cybercab went on display at a Nanjing mall on September 30, its first public showing in Jiangsu, running through October 7. The car has no steering wheel or pedals, seats two, and uses the same AI4 hardware and pure-vision stack as Tesla's consumer models. The event is a static exhibit; no sales or commercial service in China has been announced.
• Source: Nanjing Morning Post · 2026-09-30
• Editor's Take: A mall display is not a launch, but it is how a product gets normalized before regulation catches up.

2. Huawei's Qiankun ADS 5 Arrives on the Stelato V8

• Summary: Stelato V8 opened pre-sales on September 28 from ¥329,800, standard across the range with Huawei's Qiankun ADS 5 advanced driver-assist — an 896-line main lidar plus three solid-state lidars, 38 sensors in total on an L3-ready architecture.
• Source: Sina Auto · 2026-09-28
• Editor's Take: L3-ready hardware keeps shipping ahead of L3 rules. That gap is now the industry's default bet.

---

V. World Models / Physical AI

1. China's Embodied-AI Firms Reshuffle Around Interactive World Models

• Summary: A 10jqka survey argues the way Chinese embodied-intelligence firms are judged has shifted: motion demos matter less, and original world-model architecture matters more. It highlights Xingyuanzhi / 星源智's ω-EVA interactive world model, which folds prediction, verification and action into one robot control loop.
• Source: 10jqka · 2026-09-16
• Editor's Take: "Interactive world model" is the new phrase to sell. The labs already shipping one have a head start on saying it out loud.

---

物界前沿 | EastVoice