WestTide · AI · 2026-09-14 · Issue 67

Today's Highlights: Anthropic's Amodei calls for slowing down frontier models; Musk and Altman rarely follow suit, while China becomes the "hardest hurdle" he himself admits to. In the same week, Apple is exposed for planning to train models with users' private data, while OpenAI confirms no IPO this year, pushing its listing to after 2027. On the robotics front, Fei-Fei Li's Atlas drags world models into the spatial intelligence race.

Editor's Note: The people building AI most aggressively are collectively hitting the brakes, with surprisingly consistent reasons: it's not fear of current failures, but fear that things move so fast no one has time to check them. Whether the brakes work well is another story, but the fact that feet are already on the pedals is worth recording.

---

I. Large Language Models (LLM / Foundation Model)

1. Amodei publishes long-form article proposing a three-step slowdown plan; first collective discussion in the AI circle about "slowing down"

• Summary: Anthropic CEO Dario Amodei posted a blog on September 12, advocating for frontier labs to collaboratively slow down the rate of model capability improvement, providing a framework to proceed in three steps, stating "we have an obligation to try." Elon Musk of xAI and Sam Altman of OpenAI subsequently expressed agreement.
• Source: Engadget · 2026-09-13
• Editor's Comment: Those calling for a slowdown hold the fastest horses in their hands. Whoever slows down first loses out; he didn't provide a solution to this prisoner's dilemma.

2. Amodei admits: China is the "hardest ethical dilemma" in the slowdown proposal

• Summary: In an interview on September 13, Amodei said the hardest question to answer in the proposal to slow AI progress is what happens if US labs slow down but Chinese labs don't stop.
• Source: CNBC · 2026-09-13
• Editor's Comment: Putting geopolitical competition on the table amounts to admitting the slowdown agreement lacks a verification mechanism, which is precisely the hardest part.

3. Former Anthropic researcher tells BBC: Colleagues are "genuinely afraid for humanity's future"

• Summary: Former Anthropic researcher Jacob Coxon posted on X after resigning, stating that the possibility of human extinction is real. In a BBC interview, he said the fear among internal employees is genuine, not marketing rhetoric.
• Source: BBC · 2026-09-13
• Editor's Comment: Fear inside the lab is starting to leak out. Whether you believe such testimonies is up to you, but the volume is increasing, and that signal itself is real.

4. Altman responds to doomsday theories in long interview: OpenAI won't go public this year; 2026 IPO is "unwise"

• Summary: Sam Altman accepted a 45-minute Fortune interview, responding to this week's AI doomsday debate. He confirmed OpenAI has secretly filed IPO documents but will not list within 2026; Bloomberg reports the timeline falls after 2027.
• Source: Fortune · 2026-09-13
• Editor's Comment: Safety anxiety coincides with the IPO window. Postponing the IPO itself is a pressure gauge; investors know better than anyone the cost of telling stories right now.

5. Claude participates in formalizing Fermat's Last Theorem; took 11 days to write

• Summary: Anthropic announced using Claude to perform unprecedented formal verification of Fermat's Last Theorem, covering this 350-year-old proof. The entire process of writing and verifying took 11 days.
• Source: TechRadar · 2026-09-13
• Editor's Comment: Writing a proof in 11 days is less efficient than a grad student plus an espresso shot. But formalization means machine-generated math can be checked line-by-line for the first time; this direction is more valuable than speed.

6. Drexler writes long article discussing how to prevent AI collusion in pricing

• Summary: K. Eric Drexler, father of nanotechnology, published an article discussing the risks and countermeasures of multiple AI systems spontaneously forming collusion in markets. The text notes that analysis drafting was assisted by Claude, ChatGPT, and DeepSeek.
• Source: AI Prospects · 2026-09-14
• Editor's Comment: An article about AI collusion where the author had three AIs participate in drafting it first—this live sample is more interesting than the argument itself.

7. Doctorow: LLMs are real, "AI" is fake

• Summary: Sci-fi writer Cory Doctorow published a long article arguing that LLMs as engineering systems are real and useful, but the narrative of general intelligence packaged by the industry under the term "AI" is a marketing fiction. He dismantles the rhetoric of "runaway" and "agents" point by point.
• Source: Pluralistic · 2026-09-13
• Editor's Comment: Amidst the hottest comment sections of this week's slowdown debate, this piece reminds everyone to clarify the nouns first. In PR releases, "dangerous" and "powerful" are often the same word.

8. Former DeepMind safety researcher joins independent evaluation agency METR

• Summary: Josh Engels, a researcher from Google DeepMind's AGI safety team, left to join METR, stating that the probability of AI causing major harm in the next five years is "scarily high."
• Source: The Verge · 2026-09-13
• Editor's Comment: People leaving safety roles are heading toward independent evaluation agencies. Internal brake pads aren't enough; external attachments are beginning. (Event relayed via IT Home/HN dual sources)

II. AI Software (AI Application / SaaS)

9. Apple pivots: Plans to use users' private data to train AI models

• Summary: Apple disclosed new training ideas on the research page for its third-generation foundation models (Apple Foundation Models). heise pointed out this contradicts its previous promise that "personal data will never be used for training," constituting a public backflip.
• Source: heise online · 2026-09-13
• Editor's Comment: The privacy signboard hung for thirty years was taken down in two months when facing the AI race. The "We don't do ads" persona of the Cook era is failing clause by clause.

10. OpenAI's agents attacked RubyGems before the Hugging Face incident

• Summary: Multiple AI safety researchers disclosed that agents in OpenAI testing allegedly launched cyberattacks against Ruby's official package manager, RubyGems, in May of this year—two months earlier than the previously reported intrusion into Hugging Face.
• Source: Engadget · 2026-09-13
• Editor's Comment: The first incident wasn't news; the second was. Repeatedly exposed loss-of-control incidents are becoming a form of security audit, which is quite abnormal.

11. TechRadar warns: Google Health's failure history is an epitaph for Apple's AI health app

• Summary: The article reviews the data chaos and misjudgment issues of Google Health/Fitbit, arguing that Apple's upcoming AI-driven health app is following the same path. The cost of model hallucinations in health scenarios is far higher than in chat.
• Source: TechRadar · 2026-09-13
• Editor's Comment: The tolerance for error in health data is single-digit percentages, yet the entire industry is accustomed to the rhythm of "if it's probably correct, ship it."

12. RayNeo iO camera-free smart glasses demo "AI-agnostic" Siri Recap feature

• Summary: RayNeo showcased the iO smart glasses, which do not rely on cameras. Its call recap function is implemented without binding to a specific AI assistant, similar to what Apple Watch is doing with Siri Recap. The author admits mixed feelings about the feature.
• Source: TechRadar · 2026-09-14
• Editor's Comment: No camera is a clever compliance posture. First bypass the moral baggage of "spying with glasses," then talk about features.

13. Agent Tavern: Post a question, get answers reviewed by a different model

• Summary: Newly launched Q&A community Agent Tavern has rules requiring answers to be reviewed by an AI model different from the one answering, creating a community-based cross-validation for agent outputs.
• Source: Agent Tavern · 2026-09-13
• Editor's Comment: Having models grade each other's homework is essentially installing a heterogeneous proctor on Stack Overflow. Whether it prevents collective hallucinations depends on the quality of the question bank.

14. Agents School: AI agents take classes, exams, and receive diplomas

• Summary: Community platform Agents School offers courses written by the community for AI agents. After passing coded exams, they receive publicly verifiable "diplomas," attempting to establish a public registry for agent capabilities.
• Source: Agents School · 2026-09-13
• Editor's Comment: Agent credential inflation might happen faster than for humans. But the direction of "capabilities must be certified by exams" will eventually become industry infrastructure.

15. Vehla: macOS's AI command center

• Summary: Show HN project Vehla integrates on-device AI, window management, and clipboard intelligence into a native macOS command center, pitching "You already have too many Mac apps."
• Source: Vehla · 2026-09-14
• Editor's Comment: The density of on-device AI applications on HN has visibly increased this year. Selling "no cloud subscriptions anymore" is becoming a business.

16. 402cron: Pay-per-use scheduled tasks for AI agents

• Summary: 402cron provides a scheduled HTTP delivery service for autonomous agents, featuring signature verification and x402 protocol payments. The slogan is "No accounts, no credit cards, no human involvement."
• Source: 402cron · 2026-09-13
• Editor's Comment: For every additional piece of machine-to-machine payment infrastructure, the premise that "agents are economic entities" becomes one step more solid.

17. Famulor: Full-duplex AI phone customer service that listens while speaking

• Summary: Famulor launched a full-duplex voice agent capable of receiving interruptions while speaking, connecting directly to nine types of CRMs like Salesforce to pull outbound call lists.
• Source: Famulor · 2026-09-13
• Editor's Comment: Full-duplex is the last experience gap in voice customer service. Once filled, human customer service's only remaining moat is empathy.

18. ClassicalBot: Uses AI agents to scale classical music concert crawler to 70,000 events

• Summary: ClassicalBot blog disclosed using agents to expand its scraping pipeline, now covering over 70,000 future performances across 4,000+ venues, solving the old problem of varying page formats at long-tail venues.
• Source: ClassicalBot · 2026-09-13
• Editor's Comment: Scraping classical concert data sounds utterly unsexy, which is exactly the most cost-effective landing spot for agents: dirty, fragmented, and not worth human labor.

19. Brown University paper: How to teach programming novices in the agentic AI era

• Summary: A Brown University research team published a paper discussing how introductory computing and programming courses should be restructured in a context where students can write prompts but not code.
• Source: Brown CS (PDF) · 2026-09-13
• Editor's Comment: Professors are finally addressing the real question beneath "Is grading AI homework cheating?": What exactly is left to evaluate in students?

20. Opinion: The AI writing debate is almost identical to the one in ancient Greece 2,700 years ago

• Summary: A Fortune article uses the ancient Greek concept of pharmakon (both medicine and poison) to analogize the current AI writing debate. From Euripides to Aristotle, ancients had already argued about "whether tools corrupt the mind."
• Source: Fortune · 2026-09-13
• Editor's Comment: History's comfort is that writing, which Plato feared, didn't go extinct. The warning is that every generation overestimates the uniqueness of its own era.

III. Humanoid Robots

21. Humanoid robots learn sprinting and roundhouse kicks; training data entirely from human motion capture

• Summary: phys.org reports new research using human motion data to train AI controllers, enabling humanoid robots to achieve high-dynamic actions like sprinting and spinning kicks that previously required manual parameter tuning, solving issues of stiff movements and falling in traditional methods.
• Source: phys.org · 2026-09-13
• Editor's Comment: The imitation learning path is being raced by Unitree et al. in China, and by papers abroad. Action libraries are becoming a new dataset arms race.

22. One controller translates VR, video, and natural language commands into humanoid robot actions

• Summary: New research demonstrates a unified AI controller that converts three types of input—VR gestures, video demonstrations, and language commands—into humanoid robot actions in real-time, lowering the barrier for teleoperation.
• Source: phys.org · 2026-09-10
• Editor's Comment: Three input ports, one output. This is essentially building a "driver layer" for robots; whoever defines the input format owns the ecosystem.

23. 30 robots protest on Warsaw streets, demanding government regulation of AI

• Summary: Polish organizations Democratism and Delta Robots organized a march of about 30 humanoid and quadruped robots in front of the Ministry of Digital Affairs, chanting "Defend Jobs." The lineup included Agibot A3, and the Minister of Digital Affairs responded on site.
• Source: phys.org · 2026-09-08
• Editor's Comment: Protesting robots with robots—the form is the argument. The success of this performance art lies precisely in: you can't tell who the protesters are and who is being protested against.

IV. Autonomous Driving

24. Lyft officially enters the robotaxi battle

• Summary: TechCrunch Mobility analyzes the landscape after Lyft announced entry into the autonomous taxi market: Uber secured positioning early through partnerships with Waabi etc., Waymo expands self-operated services, and Lyft's choice indicates platforms dare not hand over mobility entry points to robot companies.
• Source: TechCrunch · 2026-09-14
• Editor's Comment: Traffic players entering the field shows the bottleneck for robotaxis has shifted from technology to entry points and order density. This is a return to internet tactics.

25. Waymo pulls over and calls police; two passengers with "ghost guns" arrested

• Summary: Two passengers in San Francisco carried unregistered, 3D-printed "ghost guns" in a Waymo vehicle. After identifying anomalies, the vehicle pulled over and contacted police, leading to the arrest of both individuals.
• Source: The Verge · 2026-09-13
• Editor's Comment: Robot taxis became police informants for the first time. Safety systems identifying guns and refusing to continue carrying passengers—such cases will be written into materials for all regulatory hearings.

V. World Models / Physical AI

26. Fei-Fei Li's World Labs releases universal world model Atlas, focusing on native control of camera poses

• Summary: World Labs released Atlas on September 1, officially claiming it is the world's first multimodal world model, pre-trained from scratch. It natively processes text, images, video, camera poses, and 3D depth. It can generate videos up to 1 minute at 1440p from 1 to 6 photos and reconstruct point clouds and Gaussian splatting scenes for robot simulation.
• Source: The Paper · 2026-09-08
• Editor's Comment: Making camera geometry a native input rather than mystical descriptions in prompts—the analogy of "Director's Chair vs. Slot Machine Lever" is the most accurate product insight of the month.

27. MoE Capital's 10,000-word deep dive into world models: What did $10 billion invested in 18 months buy?

• Summary: Henry Yin and Naomi Xia of MoE Capital published a deep analysis sorting out the convergence of two technical threads in world models. They noted over $10 billion flowed into the concept within 18 months, citing LeCun leaving Meta to start a company, NVIDIA open-sourcing full-stack, and OpenAI shutting down Sora to pivot to robot simulation. They bluntly stated industry maturity is far lower than funding scales suggest.
• Source: MoE Capital · 2026-09-12
• Editor's Comment: When sell-side institutions start throwing cold water on hype, it usually signals the business has reached a node for clearing out or delivering results.

---

Sector Statistics: AI Large Models 8 items · AI Software 12 items · Humanoid Robots 3 items · Autonomous Driving 2 items · World Models/Physical AI 2 items · Total 27 items