EastVoice · AI · 2026-10-01 · Issue #84

Editor's Note: China's open-weight models made their loudest move yet this week. Kimi K3 and GLM-5.3 landed inside OpenAI's own paid enterprise channel through Baseten, while Ant Group pushed Ling-3.1-flash toward a 1M-token context. On the hardware side, DeepSeek open-sourced its Huawei Ascend toolchain, and AMD closed on Fei-Fei Li's World Labs for $8.2B.

Commentary: Distribution is where this week's story sits. A strong model matters less than a model other companies are willing to ship inside their own products.

---

I. AI Foundation Models

1. Ant Group's Ling-3.1-flash Packs 560B Parameters Behind a 25B Budget

• Summary: Ant Group released Ling-3.1-flash, a sparse model with roughly 560B total parameters and about 25B activated per token, plus a 1M-token context window aimed at long documents, code, and task history.
• Source: ITHome · 2026-09-30
• Editor's Take: Sparse designs are how Chinese labs keep inference costs sane while still posting frontier-scale numbers on paper.

2. Chinese Open Models Enter OpenAI's Paid Enterprise Lane

• Summary: US AI infrastructure firm Baseten said it is working with OpenAI as one of the first open-model inference providers on the B2B marketplace, letting enterprises call Kimi K3 and GLM 5.3 through Codex.
• Source: ITHome · 2026-09-30
• Editor's Take: Chinese weights are now being sold inside an American lab's own storefront. That is a distribution win no benchmark score can match.

3. DeepSeek Open-Sources Its Huawei Ascend Toolchain

• Summary: DeepSeek published infrastructure components for Huawei's Ascend platform, covering a TileLang compiler, compute libraries, and a distributed communication library, each mirroring an earlier CUDA build.
• Source: ITHome · 2026-09-30
• Editor's Take: Silicon is the easy headline. The harder, slower win is the software layer that decides whether that silicon is actually usable.

4. Anthropic's Complaint About GLM-5.3 Reads Like an Ad for Zhipu

• Summary: QbitAI notes that Anthropic's warnings about GLM-5.3's software-exploitation skills have, in effect, advertised the model, with Chinese outlets reporting it matches strong closed models on several tasks.
• Source: QbitAI · 2026-09-30
• Editor's Take: Calling a rival dangerous is a strong compliment in this market. Accelerationists on both sides will read it as a spec sheet.

5. AMD Buys Fei-Fei Li's World Labs for $8.2B

• Summary: AMD announced an all-stock deal worth about $8.2B for World Labs, the spatial-intelligence startup Fei-Fei Li founded two years ago. Li joins as an executive vice president and chief scientist.
• Source: Huxiu · 2026-09-30
• Editor's Take: AMD just bought research talent at roughly $117M a head. It is a bet that world models, not raw FLOPS, become the next hardware moat.

6. OpenAI DevDay 2026: Dots, a $500 Pro Tier, and 20-Plus Updates

• Summary: At its DevDay event OpenAI unveiled Dots, an always-on agent that runs on a separate cloud computer with its own browser, alongside a $500-a-month Pro plan and a long list of smaller releases.
• Source: SSPAI · 2026-09-30
• Editor's Take: The blobby avatar is cute. The real product is a persistent agent with its own identity and network access, and that is a much bigger commitment for users.

7. OpenAI Puts GPT-6.1 Sol in Codex at a Lower Price

• Summary: OpenAI rolled out GPT-6.1 Sol across Codex and ChatGPT Work, describing performance close to its top-tier Astra model at meaningfully lower cost.
• Source: ITHome · 2026-09-30
• Editor's Take: Cheap-and-nearly-as-good is the number that moves enterprise adoption. Frontier bragging rights rarely close a procurement deal.

8. Apple Adds Web-Aware "Use Model" to iOS 27 Shortcuts

• Summary: Apple updated support documents for iOS 27, iPadOS 27, and macOS 27, adding over 35 Shortcuts actions, led by a Use Model command that can reach the web for current information.
• Source: ITHome · 2026-09-30
• Editor's Take: Siri has been the punchline for years. Quietly bolting real model access onto Shortcuts may end up mattering far more than a new voice.

9. Microsoft's Project Quine Treats Biology as a World Model

• Summary: Microsoft Research introduced Project Quine, an experimental multimodal system billed as a world model for biology, connecting computational modeling with real wet-lab experiments, in work with Harvard.
• Source: ITHome · 2026-09-30
• Editor's Take: Drug discovery is where world models can claim a payoff beyond a demo video. It is also where being wrong is expensive.

10. ByteDance's Doubao and ChatGPT Bleed Cash for Different Reasons

• Summary: Huxiu runs the numbers on how ByteDance's Doubao and OpenAI's ChatGPT both lose money on free usage, and how Doubao's commercial team was reportedly cut to a third as it reorganizes around monetization.
• Source: Huxiu · 2026-09-30
• Editor's Take: Free scale and paid scale are different businesses. ByteDance is discovering that millions of users can be a liability before they are a moat.

---

II. AI Software & Applications

1. Tencent Builds Three "Muse" Products and a Personal Agent

• Summary: Tencent is reportedly developing a personal agent codenamed Handy Bot as a standalone app, part of a push that has produced three products under the Muse name.
• Source: Huxiu · 2026-09-30
• Editor's Take: Tencent rarely moves first, but it moves everywhere at once. Three shots at the same target suggest it is not sure which shape wins.

2. Five Agent Products in Two Months, All Chasing "Long-Term Delegation"

• Summary: A Huxiu roundup counts five major agent launches in two months, including OpenAI's DevDay releases, arguing the race has shifted from single tasks to handing an agent sustained responsibility.
• Source: Huxiu · 2026-09-30
• Editor's Take: "Long-term delegation" is a polite term for letting software act while you are not watching. Nobody has priced that trust yet.

3. Robinhood Ships an AI Agent That Trades for You

• Summary: Robinhood unveiled a suite of new tools at its annual event, including an AI agent that can watch the market and execute trades on a user's behalf.
• Source: ITHome · 2026-09-30
• Editor's Take: The CEO says the agent apps are safe. When the position is down, "safe" is not the word users will reach for.

4. OpenAI in Talks to Raise $30B Pre-IPO at a $1.4T Valuation

• Summary: OpenAI is negotiating a pre-IPO round of at least $30B at a valuation near $1.4 trillion, according to reports, months after its compute commitments ballooned.
• Source: ITHome · 2026-09-30
• Editor's Take: A trillion-dollar private valuation takes two sides to agree. The harder sell is to the public market later, once the numbers are audited.

5. Apple Shelves Plan to Replace About 5,000 AppleCare Staff With AI

• Summary: Bloomberg's Mark Gurman reports Apple had considered cutting roughly 5,000 AppleCare support roles in favor of AI but has decided to hold off.
• Source: ITHome · 2026-09-30
• Editor's Take: Deploying AI customer service is easy to announce and hard to survive. Apple blinking here says more than any roadmap slide.

6. Anthropic Signs Compute Deal With SpaceX Worth Up to $84.5B

• Summary: Filings tied to SpaceX's IPO reveal a compute agreement with Anthropic reportedly worth up to $84.5B, cancellable with 90 days' notice, as Anthropic prepares its own listing.
• Source: ITHome · 2026-09-30
• Editor's Take: An anchor tenant with a 90-day exit clause is not a handshake. It is a reminder of how thin these mega-commitments can be.

7. AI Travel Assistants Stumble Once They Meet Real Trips

• Summary: A Huxiu field test of AI travel helpers found the promised one-click planning falls apart on real bookings, from ticket timing to cross-app price comparison.
• Source: Huxiu · 2026-09-30
• Editor's Take: Holiday travel is a brutal benchmark. It rewards agents that can actually finish a transaction, not ones that produce a nice itinerary.

8. OpenAI's Real Target Looks a Lot Like "WeChat"

• Summary: GeekPark argues OpenAI's DevDay releases, from shared spaces to collaboration tools, point at a super-app ambition, even as subscription quotas get cut for some tiers.
• Source: GeekPark · 2026-09-30
• Editor's Take: The WeChat comparison flatters OpenAI. The company still has to build the payment rails, habits, and stickiness that took Tencent a decade.

---

III. Humanoid Robots

1. 36 Listed Robot Makers, and a Brutal Gap in Profitability

• Summary: A Huxiu survey of 36 robot companies that have gone public finds earnings diverging sharply, with Unitree, Mech-Mind, and others newly listed and the market now scrutinizing real commercialization.
• Source: Huxiu · 2026-09-30
• Editor's Take: Once the bell stops ringing, the only question left is margin. The IPO wave is about to become a sorting machine.

2. Embodied AI's Body-Less Data Finally Has a Flagship Result

• Summary: Leiphone reports a result for the "body-less" data problem in embodied learning, where a robot completes tasks without relying on data collected from its own specific hardware.
• Source: Leiphone · 2026-09-30
• Editor's Take: Data that transfers across bodies is what turns scarce robot footage into a real asset. It is the least glamorous and most decisive bottleneck.

3. Why Vision-Tactile Sensing Is Hard to Copy

• Summary: Leiphone examines vision-tactile sensing, the technology that lets a robot seat a USB cable into a millimeter-scale port, and explains why it resists easy replication despite looking simple.
• Source: Leiphone · 2026-09-30
• Editor's Take: Assembly lines pay for precision, not for demos. Whoever solves the fingertip owns the 3C factory floor.

4. Home Robots Face a Fork: Companion or Worker?

• Summary: TMTPost compares Figure's Helix 2.5, deployed across 30 homes with no price published, against Qiyuan's consumer robots priced from ¥19,999 to ¥29,999, and asks what buyers actually want.
• Source: TMTPost · 2026-09-30
• Editor's Take: Figure is selling capability, Qiyuan is selling a shopping decision. The one that names a price is the one closer to a business.

---

IV. Autonomous Driving

1. Huawei Tells Automakers to Stop Doing Everything In-House

• Summary: Huawei's senior vice president Jin Yuzhi argued automakers should not build every intelligent-driving system themselves, pressing the case that sustained R&D only pays off at scale.
• Source: Huxiu · 2026-09-30
• Editor's Take: This is Huawei arguing for Huawei, but the math is real. A single-brand software stack is an expensive way to learn the same lessons.

2. Self-Driving Violations and the Problem of Self-Policed Evidence

• Summary: A legal-tech essay in Huxiu flags a weak point in China's draft rules: when an autonomous car runs a red light, the evidence comes from the automaker's own data.
• Source: Huxiu · 2026-09-30
• Editor's Take: Asking the accused to grade their own paper is not a small detail. It is the difference between a regulation and a press release.

3. Tesla's FSD (Supervised) Wins Approval in Croatia, Europe Count Hits Eight

• Summary: Croatia approved Tesla's FSD (Supervised) under national rules, making it the eighth European country to allow the system while the EU's bloc-wide vote stays stalled.
• Source: ITHome · 2026-09-30
• Editor's Take: Country-by-country approval is slow and unglamorous, but it is how Tesla gets a European footprint while Brussels argues with itself.

4. Tesla Signs a $30B Credit Line to Fund AI Compute and Solar

• Summary: A regulatory filing shows Tesla secured $30B in credit agreements, including a $20B delayed-draw term loan, to expand AI compute and solar capacity.
• Source: ITHome · 2026-09-30
• Editor's Take: The balance sheet is being stretched for compute and robots at once. Both bets pay off, or neither does.

---

V. World Models / Physical AI

1. Embodied AI Leaves the Ground: Drones With Robot Arms (IROS 2026)

• Summary: Leiphone covers IROS 2026 work that pushes embodied intelligence from ground-based manipulation toward aerial cooperation, mounting robotic arms on drones for tasks such as tightening bolts.
• Source: Leiphone · 2026-09-30
• Editor's Take: Bolting an arm onto a drone is a small mechanical change and a large control problem. Whoever solves it opens a category with no incumbent.

---

物界前沿 | EastVoice