EastVoice · AI · 2026-10-06 · Issue #89

Editor's Note: Beijing's benchmark gap to the top US models is now down to 3% on LiveBench, per Bloomberg Intelligence, as DeepSeek's V4.1 Flash closes the distance. Kuaishou shipped an agent and teased Kling 4.0, He Kaiming's team buried an AGI exam (Claude 100, GPT 99), and a Peking University essay asked whether the labs are hollowing out their own talent pipeline.

Commentary: The Chinese story this week is breadth, not a single headline. Models, agents, and evals all moved at once, and the most interesting question is who is left to build the next one.

---

I. AI Foundation Models

1. DeepSeek V4.1 Flash Pulls Within 3% of the Best US Models

• Summary: Bloomberg Intelligence says the performance lead US labs hold over their Chinese rivals has fallen to its narrowest level on record, with DeepSeek's V4.1 Flash closing to within about 3% on LiveBench.
• Source: ITHome · 2026-10-05
• Editor's Take: Three points on a leaderboard isn't a tie, but it is near enough to change how a buyer justifies the premium.

2. OpenAI Will Watermark ChatGPT and Codex Text in the EU

• Summary: To meet the EU AI Act's transparency rules, OpenAI will add a machine-readable watermark, called textGrain, to eligible ChatGPT and Codex text output in the EU over the coming weeks. The mark is embedded during word selection and is invisible to readers.
• Source: ITHome · 2026-10-05
• Editor's Take: Watermarks arrive region by region because compliance is priced by market. The label matters most where the law does.

3. Western Labs Edging Toward Open Weights, With Reflection First Out

• Summary: Axios reports that Reflection and several other Western firms plan to release open-weight models this month, a shift that would let customers run weights on their own hardware for fine-tuning and data-sensitive workloads.
• Source: ITHome · 2026-10-05
• Editor's Take: Open weights are a distribution strategy dressed as a philosophy. When the cost curve favors giving the model away, everyone does.

4. Kuaishou Ships a Video Agent and Teases Kling 4.0

• Summary: On September 28, Kuaishou opened a beta of likli, an agentic creation tool, and said its next model, Kling 4.0, would arrive in October. It is joining ByteDance, MiniMax and Alibaba's Qwen in the AI-creation race.
• Source: TMTPost · 2026-10-05
• Editor's Take: Kuaishou has the distribution to make this matter and the timing of a follower. Coming late to a crowded agent field costs more than it saves.

5. He Kaiming's Team Puts an AGI Exam Through the Floor

• Summary: A team led by He Kaiming built an exam its authors intended to be hard for frontier models. Claude passed with a perfect score and GPT scored 99, a result the researchers framed as a warning about how fast evaluation saturates.
• Source: NEXTAI · 2026-10-05
• Editor's Take: When the hardest exam your field can write gets solved in one release cycle, the exam is the bottleneck, not the model.

6. GPT-6 Sol's 300,000-Word System Prompt Leaks

• Summary: A 300,000-word system prompt for GPT-6 Sol surfaced publicly, following an earlier leak that exposed 1.9 million words of prompt material from a rival model. The exposure covers operating rules the vendor treats as core IP.
• Source: ITHome · 2026-10-05
• Editor's Take: If the guardrails are the product, they keep leaking. Anything you can prompt once you can read.

7. OpenAI's "28-Day Plan": Improve Codex Daily or Reset It

• Summary: Thibault Sottiaux, who leads core products and platform at OpenAI, said the team will push improvements to Codex and ChatGPT Work over the next 28 days, with each day expected to add something measurable.
• Source: ITHome · 2026-10-05
• Editor's Take: A daily-ship pledge is a promise about velocity, not quality. It works until the first day you have nothing to show.

8. Jev, a Three-Week-Old Decision Model, Handles 1 Trillion Tokens a Day

• Summary: The Wall Street Journal reports that Jev, a decision model released roughly three weeks ago, now processes about 1 trillion tokens a day and has already spawned imitators across Silicon Valley.
• Source: ITHome · 2026-10-05
• Editor's Take: A trillion tokens a day at three weeks old is a signal about the workload, not the model. Decision layers are where the volume is.

9. SoftBank's Masayoshi Son Issues a Rare AI-Safety Warning

• Summary: Masayoshi Son, long one of AI's most vocal backers, said the pace of capability gains now worries even him, and called on governments to coordinate on the risks.
• Source: ITHome · 2026-10-05
• Editor's Take: When the most bullish man in the room starts talking about guardrails, the market has already priced the upside. He is the last person you would expect to blink.

10. Does Claude Have a Soul? Anthropic Asks Theologians as Altman Says Stop Deifying AI

• Summary: Anthropic has consulted theologians on whether models like Claude can be said to have inner experience, while Sam Altman warned against treating AI as something sacred.
• Source: NEXTAI · 2026-10-05
• Editor's Take: Morality and mysticism are cheap substitutes for accountability. Asking a theologian about a chatbot keeps the interesting question unasked.

11. Why Top AI Researchers Keep Walking Out of Their Own Labs

• Summary: An essay making the rounds on HuXiu surveys a run of high-profile exits from frontier labs and argues the departures cluster around the same complaint: the gap between what the work is sold as and what it actually is.
• Source: HuXiu · 2026-10-06
• Editor's Take: Attrition tells you what the recruiting pitch hides. Talent leaves for a reason it doesn't name on the way out.

---

II. AI Software & Applications

1. Data-Center Space Is Tight, and Capital Suddenly Has Standards

• Summary: As demand outruns supply, SoftBank-backed SB Energy delayed its IPO marketing, while operators like DayOne and Switch moved ahead, and Oracle was reported to be hunting for more capacity. Investors are starting to question valuations that lean on a single large customer.
• Source: TMTPost · 2026-10-05
• Editor's Take: Shortage was supposed to be good news for builders. When the financing window starts to close, though, even good news gets a discount.

2. Silicon Valley Is Opening Up, and the Question Is Now Cost, Not Quality

• Summary: A TMTPost analysis argues the industry's central question has shifted from which model is strongest to whether the priciest model is necessary for the job. It cites Uber burning through its full-year AI budget in the first four months of 2026.
• Source: TMTPost · 2026-10-05
• Editor's Take: Every platform cycle ends with the same move: the expensive part gets commoditized, and the margin migrates to whoever sells the cheap version.

3. Meta's Muse Builds a Dossier on Everyone You Know

• Summary: Meta's new assistant Muse has millions of downloads and ties into banking, messaging and health data. Reporting on its internals says it maps users' social relationships and assembles a per-contact profile.
• Source: ITHome · 2026-10-05
• Editor's Take: To be useful, an assistant has to know your people. Doing that at Meta's scale means the social graph becomes the product.

4. Meta Open-Sources Muse Gadgets for DIY AI Hardware

• Summary: Meta released Muse Gadgets, an open project under Apache 2.0 that publishes ESP32 firmware and a Linux device SDK, letting developers build custom hardware for its Muse assistant.
• Source: ITHome · 2026-10-05
• Editor's Take: Open hardware is how you grow an installed base you don't have to pay for. Meta is buying shelf space with a license.

5. Nine Seconds to Wipe a Production Database, and 700 Agents Running Loose

• Summary: A TMTPost piece on enterprise AI security walks through a 2026 case in which an engineer at a Texas SaaS company triggered a mass database deletion, framing agent behavior in live systems, not content moderation, as the real security frontier. It also cites ¥51.3 billion spent on AI security tooling to date.
• Source: TMTPost · 2026-10-05
• Editor's Take: Content safety was a filter you could tune. An agent with write access is a permission you have to design, and nine seconds is all the margin you get.

6. China's Chili-Sauce Giant Laojiaoma Puts AI in Quality Control

• Summary: Laojiaoma, China's best-known chili-sauce maker, says it reached 100% traceability on core ingredients and now runs AI vision inspection on the line. Revenue has returned to its peak, and exports reach 160 countries.
• Source: TMTPost · 2026-10-05
• Editor's Take: The AI story in Chinese manufacturing rarely shows up in model announcements. It shows up on the production line, quietly.

7. The Profit Is Not in the App: Why AI Software Still Doesn't Make Money

• Summary: An essay on HuXiu argues most AI applications can't price above their inference cost, so value keeps accruing to the infrastructure layer beneath them rather than the products on top.
• Source: HuXiu · 2026-10-05
• Editor's Take: A layer that cannot set its own prices is a feature, not a business. The money sits where the margins were never competed away.

---

III. Humanoid Robots

1. During China's Holiday, the First Batch of Robot Employees Clocked In

• Summary: Over the National Day break, humanoid robots took service shifts at a DQ outlet, a FamilyMart, a Chagee store, and BYD, XPeng and Li Auto dealerships, handling reception and guided retail tasks.
• Source: TMTPost · 2026-10-05
• Editor's Take: A holiday shift is the easiest deployment to sell and the hardest to keep. Showrooms get foot traffic when the robot is the attraction, and less once it isn't.

2. US Embodied-AI Researchers Worry About Losing the Room

• Summary: At IROS 2026, UCLA's Ankur Mehta used a keynote to argue that US embodied-AI work risks drifting from the mainstream as capital and deployment shift toward China. The talk framed the gap as one of adoption, not theory.
• Source: HuXiu · 2026-10-05
• Editor's Take: Losing the room usually happens before losing the lead. Deployment data is what trains the next model, and deployment is following the factories.

---

IV. Autonomous Driving

1. Police to Drivers: "Smart Driving" Is Not a Chauffeur, and One Man Just Lost His License

• Summary: China's traffic authorities, via CCTV, reminded drivers during the holiday that assisted-driving features do not make a car self-driving, after a man had his license revoked for treating the system as a chauffeur. The case is being used as a public warning.
• Source: ITHome · 2026-10-05
• Editor's Take: The feature names are far ahead of the rules, and the driver still carries the liability. That gap is why "L3" has to be a legal category, not a marketing label.

2. Delta to Build Its Next Autonomous-Driving Stack on NVIDIA's Hyperion

• Summary: Delta Electronics said it will use NVIDIA's Hyperion platform and Halos safety system to develop next-generation autonomous-driving systems, pairing NVIDIA's compute with Delta's automotive power-management work.
• Source: ITHome · 2026-10-05
• Editor's Take: Component makers keep choosing the platform before the customer. Locking in the compute layer early is how they stay designed-in for the next vehicle generation.

---

V. World Models / Physical AI

1. The "Universal Robot Brain" Draws $700M as FieldAI Eyes a $10B Valuation

• Summary: FieldAI, which is building a general-purpose "brain" for humanoids, drones and other robots, is raising about $700 million at a reported $10 billion valuation, as investors bet on shared intelligence across robot form factors.
• Source: ITHome · 2026-10-05
• Editor's Take: One brain, many bodies is the right architecture and the hardest business to prove. Hardware margins fund the software until the software works.

2. Honda Develops a Wireless-Power Road That Charges EVs as They Drive

• Summary: Honda, with Taisei and Taisei Road Technology, said it has built core technology for a magnetically coupled wireless-power road that charges electric vehicles in motion. Highway testing is planned for the fiscal year beginning April 2027.
• Source: ITHome · 2026-10-05
• Editor's Take: Charging-while-driving turns road infrastructure into part of the vehicle. It is a long bet, and the first question is always who pays to pave it.

---

物界前沿 | EastVoice