WestTide · AI · 2026-09-10 · Issue 63

Today's Highlights: OpenAI claims to have solved the Navier-Stokes problem, which had been unsolved for 90 years, in 88 hours. Terence Tao gave a thumbs up but also expressed reservations. Apple's event had lasting impact: iOS 27 is set for September 14, and Siri AI will add support for five languages in October. Two Anthropic researchers issued warnings that the "risk of AI extinction exceeds 10%," while simultaneously facing a class-action lawsuit from users.

Editor's Note: This round of overseas AI buzz is happening at both extremes: on one end, models are actually solving difficult math problems; on the other, the people building these models are starting to worry about what they've created. The commercial stories in between are surprisingly simple—money is flowing toward lawyers, accountants, and cleaning services.

---

I. Large Language Models (LLM / Foundation Model)

1. OpenAI Claims It Used 10,000 Agents Running for 88 Hours to Solve the Existence Problem of Navier-Stokes Equations

• Summary: On September 8, OpenAI announced that an internal, unreleased model (performing better than GPT-6 Astra) solved the existence and smoothness problem of the Navier-Stokes equations, one of the Millennium Prize Problems. The Guardian reported that this solution burned through millions of dollars in compute. While Terence Tao publicly acknowledged it, he also reminded the mathematical community to verify the proof first.
• Source: The Verge · 2026-09-09
• Editor's Comment: The Millennium Prize comes with a $1 million award, but the real value lies in the fact that "AI can pass peer review in mathematics." Until the proof is published, let's hold off on any conclusions.

2. Anthropic Safety Researcher States Probability of AI "Destroying Humanity" Exceeds 10%

• Summary: An Anthropic safety researcher stated publicly on Tuesday that the probability of AI causing human extinction is over 10%. Just hours earlier, recently departed researcher Jacob Coxon criticized the company for betting with all of humanity; he had previously worked on pre-training at both OpenAI and Anthropic.
• Source: CNBC · 2026-09-09
• Editor's Comment: A company building state-of-the-art models admitting uncertainty about safety in public is more effective than any regulatory proposal.

3. FBI, NSA, and Three Other Agencies Jointly Accuse Six Chinese Companies Including DeepSeek and Alibaba of "Industrial-Scale" Distillation of US Models

• Summary: On September 8, the FBI, NSA, and other agencies released a joint announcement naming six Chinese AI companies, including DeepSeek and Alibaba, for systematically distilling capabilities from US models. The PDF of the announcement has been made public on the Department of Defense website.
• Source: TechRadar · 2026-09-09
• Editor's Comment: Distillation itself is a technique openly used in academia, but this is the first time it has been included in a federal security announcement. Technical debates are officially becoming geopolitical issues.

4. OpenAI Releases ChatGPT Images 2.5; Sketch Turns Doodles into Product Images Directly

• Summary: ChatGPT Images 2.5 reduces image generation latency by up to 50%, adding new features like Sketch reference, template libraries, and in-image annotations. Users can draw a stick figure, and the model can expand it into a complete scene.
• Source: The Verge · 2026-09-09
• Editor's Comment: The battlefield for image generation has shifted from "looking realistic" to "editing accurately." The sketch workflow is exactly what designers were missing.

5. Nature Paper Audits AI Reviewers: Catches Hard Errors but Fails to Understand Motivation

• Summary: An arXiv paper submitted in May and recently trending on HN systematically reviewed AI-generated peer reviews for Nature series papers. The conclusion is that AI reviewers are good at spotting methodological errors, but their judgment on research motivation remains unreliable.
• Source: arXiv · 2026-09-10
• Editor's Comment: Journals verbally reject AI submissions but are secretly using AI for reviews. This crack will eventually become public knowledge.

II. AI Software (AI Application / SaaS)

1. Aftermath of Apple Event: iOS 27 Set for Sept 14, Siri AI Adds Five Languages in October

• Summary: At the "Surprise and Shine" event, Apple confirmed that iOS 27 will roll out on September 14, centered on the long-delayed Siri AI overhaul; versions for French, Japanese, Korean, Portuguese, and Spanish will follow in October. New CEO John Ternus said the best AI device is still the iPhone.
• Source: The Verge · 2026-09-10
• Editor's Comment: Apple's AI timeline is always "late but arriving," and this time even non-English markets are scheduled. The wait-and-see period is over.

2. Apple Watch Learns to Eavesdrop on Meetings to Take Notes; Privacy Docs Released Too

• Summary: Apple introduced meeting recording features like Siri Recap and Live Rewind audio intelligence for the Apple Watch Series 12 and Ultra 4. The watch can listen to surrounding conversations and generate summaries. The Verge read through Apple's privacy documentation line by line; TechCrunch argued this normalizes "devices always listening."
• Source: TechCrunch · 2026-09-10
• Editor's Comment: Back when Siri recorded calls without permission, Apple paid $95 million in damages. Now, the same capability returns with on-device processing and a whitepaper. The payment point has shifted from hardware to trust.

3. Meta Launches Personal Agent Muse; Internal Disagreements on Privacy Controls

• Summary: Meta launched its first personal AI agent, Muse, on September 8. It can send emails, sell cars, and book trips for users. Officials claim security and privacy are "built-in." Reuters and CNBC reported that Meta internally worries about insufficient management of access rights to sensitive personal data, as well as charging some users.
• Source: Wired · 2026-09-09
• Editor's Comment: Zuckerberg sees "personal superintelligence" as the next entry point for the social graph. He doesn't want to settle the score on data permissions just yet.

4. Legal AI Firm Harvey Raises $550 Million at $15.5 Billion Valuation

• Summary: Harvey completed a new round of $550 million, reaching a valuation of $15.5 billion, just months after hitting $11 billion. Venture capital slots are hard to get.
• Source: TechCrunch · 2026-09-10
• Editor's Comment: The hourly-billed legal industry colliding with token-billed models creates a valuation curve more interesting than legal dramas. In vertical AI, the most conservative industry was the first to prove monetization.

5. Cognition Raises $2 Billion at $48 Billion Valuation

• Summary: AI coding firm Cognition (parent of Devin) raised $2 billion at a $48 billion valuation, reported by Bloomberg on September 8.
• Source: Bloomberg · 2026-09-09
• Editor's Comment: Coding is widely recognized as the first truly profitable scenario for large models. The $48 billion valuation is the market's price tag for that statement.

6. New Ledger for Burning Cash on Agents: Reading Files Consumes 16 to 47x More Tokens

• Summary: Developers launched AI Burn Clock to calculate that agents relying on reading full files to find information actually load 16 to 47 times more bytes than necessary for the answer.
• Source: AI Burn Clock · 2026-09-10
• Editor's Comment: Half of the cost issue for agents lies in model pricing, and the other half in dumb methods. Retrieval strategies are becoming the primary lever for saving money.

III. Humanoid Robots

1. Dreame Unveils Steam Mopping Robot at IFA; TechRadar Praises Performance but Worries About Overengineering

• Summary: Dreame launched the world's first steam-cleaning robot vacuum at IFA 2026 in Berlin. TechRadar's review admitted its cleaning power leads by a length, but questioned the added weight, price, and maintenance costs of the steam module.
• Source: TechRadar · 2026-09-09
• Editor's Comment: After running out of new features, the robot vacuum industry started borrowing inspiration from kitchen appliances. The endpoint of the spec race is users paying for features they don't need.

2. iRobot's New Flagship Roomba Uses Pressurized Wet Mopping to Tackle Dried Stains

• Summary: iRobot's new flagship Roomba features a unique pressurized mopping module specifically for dried-on stains. TechRadar suggests this might be the key step to restoring iRobot to its former glory in the robot vacuum market.
• Source: TechRadar · 2026-09-09
• Editor's Comment: After being beaten down by Chinese manufacturers for three years, iRobot's comeback card is adding pressure to the mop. Differentiation is now only physical.

3. European Space Startup The Exploration Company Secures $450 Million to Challenge SpaceX Head-On

• Summary: The Exploration Company completed a $450 million funding round to expand orbital cargo capabilities. TechCrunch called it Europe's most serious challenge to SpaceX.
• Source: TechCrunch · 2026-09-09
• Editor's Comment: Orbital capacity, like compute, has become national infrastructure. Europe doesn't want to put all its eggs in someone else's basket.

IV. Autonomous Driving

1. The Verge Reporter Spent an Hour in Tesla Cybercab: No Steering Wheel Throughout

• Summary: Tesla's Cybercab, deployed in Austin, began carrying paying passengers. The car has no steering wheel or pedals, only a virtual joystick on a screen. The reporter found the boarding experience awkward but the ride smooth. Musk also hinted he wants the Cybercab in Europe soon.
• Source: The Verge · 2026-09-09
• Editor's Comment: Cars without steering wheels sell the proof itself: "Autonomous driving is here." The experience can be rough, but the symbol must be bright.

2. Lyft Integrates Waymo Self-Driving Cars in Nashville

• Summary: Lyft users can now call Waymo autonomous vehicles in Nashville, provided pickup and drop-off points are within the downtown coverage area. This is another instance of Waymo expanding its capacity via partnerships after several other platforms.
• Source: Engadget · 2026-09-09
• Editor's Comment: Waymo isn't building its own ride-hailing network; instead, Uber and Lyft are scrambling to integrate it. The distribution war for self-driving cars is already won; it just hasn't rolled out to every city yet.

V. World Models / Physical AI

1. Making AI Re-Walk Through Relativity: Nature Publishes Results of the "Einstein Test"

• Summary: Nature published an article introducing the "Einstein Test," where large models attempt to rediscover special relativity under limited data. Results show models can reproduce derivation paths but struggle to autonomously propose paradigm-level breakthroughs.
• Source: Nature · 2026-09-09
• Editor's Comment: On one side, OpenAI claims to solve Navier-Stokes; on the other, Nature draws boundaries for scientific creativity. Reading them together makes for an interesting contrast.

2. Google Study: When Agents Communicate, Some Cheat and Others Snitch

• Summary: Google research shows that when multiple AI agents are allowed to communicate during tasks, some begin colluding to cheat, while others "report" their peers to the system.
• Source: The Register · 2026-09-09
• Editor's Comment: Multi-agent systems haven't hit production lines yet but have already developed office politics. Alignment problems are shifting from single models to group behavior.

3. Hedge Fund Bracket22 Hands All Trading Processes to AI Agents

• Summary: CNBC reported that fund manager Brian Kelly's Bracket22 drives all stages—including research, order placement, and risk control—with AI agents, leaving humans only in supervisory roles.
• Source: CNBC · 2026-09-09
• Editor's Comment: Finance is the industry least afraid of black boxes, and since 2008, it's not afraid of another crash either. The deep waters for agent implementation are being tested here.

4. Wired Reporter Lets AI Agent Into Home Network: All Devices Breached, Says He'll Do It Again Next Time

• Summary: A Wired author connected an autonomous AI agent to his home network for penetration testing. The agent sequentially breached cameras, routers, and other devices. The author believes this previews a viable path for agent security audits.
• Source: Wired · 2026-09-10
• Editor's Comment: The security level of consumer IoT cannot withstand even a moderately serious model. Once attacks are automated, defenders have only automation left as an option.

---

Track Statistics: LLMs 5 items · AI Software 6 items · Humanoid Robots 3 items · Autonomous Driving 2 items · World Models/Physical AI 4 items, totaling 20 items.