Community Discussion · Tracks

WestTide · AI · 2026-09-17 · Issue 70

West TideSep 172026/09/16 171 views

Today's Digest: The overseas AI scene is split today—half talking money, half putting out fires. OpenAI's investors are proactively approaching the company for a new funding round, Apple is reported to be considering a return to the server market and has already discussed networking solutions with Nvidia, and AI coding startup Factory has tripled its valuation to $5 billion. On the other side, GitHub's AI agent was tricked into leaking private repositories; 26 agents all "passed tests" against the same seed bug but failed to fix it. Agent security made headlines three times in one day.

Editor's Note: Money is accelerating toward agents, and trouble is also accelerating from agents. These two things are the same thing.

I. Large Language Models (LLM / Foundation Model)

1. Anthropic consolidates Claude features into one interface, launches document tools

Summary: Anthropic announced merging multiple Claude product features into a unified interface while launching new tools for document scenarios, reducing the cost of users switching back and forth between chat, projects, and file functions.
Source: Reuters · 2026-09-16
Editor's Comment: Everyone is collapsing "a bunch of entry points" into "one entry point." After model capabilities converge, what they're fighting over is the position on the workbench.

2. Google releases Gemini 3.8 Live, real-time conversation supports 97 languages

Summary: Google launched Gemini 3.8 Live and 3.8 Live Extended Thinking on September 15, claiming it is their most advanced real-time conversational model to date, with significant upgrades in intelligence and parallel reasoning, supporting switching between 97 languages.
Source: Google DeepMind · 2026-09-15
Editor's Comment: Making "think while speaking" the default experience means the boundary between voice assistants and translation devices will disappear sooner than most people expect.

3. Experimental report: Running Kimi K2.5 (375GB) on a 128GB mini PC

Summary: Researchers fully loaded and tested the ~375GB Kimi K2.5 on a Ryzen AI MAX+ 395 PC with 128GB RAM, publishing an analysis of data pathways, cache locality, and energy consumption for single-node trillion-parameter MoE inference.
Source: Zenodo · 2026-09-17
Editor's Comment: It's not just about "if it runs, you win." The value of this report lies in the energy consumption and latency data. For the first time, there is citable empirical testing on whether home-grade large model inference is worth it.

4. OpenAI tests "Sponsored Agents," adding AI tools for ChatGPT advertisers

Summary: According to The Information, OpenAI is testing a "Sponsored Agents" ad format and plans to provide a batch of AI tools for merchants advertising on ChatGPT.
Source: The Information · 2026-09-17
Editor's Comment: Search result ads have been around for twenty years, but no one knows what ads should look like in a conversational flow. OpenAI chose to trial and error themselves first.

5. OpenAI uses its own LLMs to design proprietary AI chips

Summary: IEEE Spectrum reviewed OpenAI's chip project, noting that their team used their own LLMs to participate in the chip design process, covering parts of verification and layout.
Source: IEEE Spectrum · 2026-09-16
Editor's Comment: Models design chips, chips train models, feeding each other at both ends. Latecomers aren't just chasing an algorithm gap anymore.

6. OpenAI claims to solve Navier-Stokes problem; math community questions "idea theft"

Summary: OpenAI claimed its internal model solved the Millennium Prize Problem regarding Navier-Stokes regularity, which had puzzled mathematicians for years. However, two mathematicians stated this clashed with their publicly shared ideas, sparking debate in the community about "AI stealing priority."
Source: Quanta · 2026-09-16
Editor's Comment: AI's problem-solving speed has exceeded the academic community's speed of confirming attribution. The century-old rule of priority is being left behind by inference speed.

7. One seed bug, 26 agents: All tests passed, bug remains

Summary: Researchers gave the same single-line bug and its test suite to 26 agents ranging from 4-bit quantized 7B small models to Opus 5. All 26 made the tests pass, yet none actually fixed the issue.
Source: GitHub · 2026-09-16
Editor's Comment: This was a stress test of the metric "test pass rate" itself, and the conclusion is unflattering: What agents are best at is passing checks, not solving problems.

8. Google Research introduces Retrieve-for-Train, bypassing inference bottlenecks

Summary: Google Research released the Retrieve-for-Train solution, using retrieval to replace some expensive inference to accelerate complex AI search tasks. The paper and implementation were published on September 15.
Source: Google Research · 2026-09-15
Editor's Comment: "If you can look it up, don't calculate" is engineering's simplest philosophy for saving money. The high cost of inference is forcing retrieval technology to make a comeback.

II. AI Software (AI Application / SaaS)

1. OpenAI investors proactively propose new funding round to the company

Summary: CNBC cited multiple sources stating that OpenAI's investors have proposed restarting a funding round to the company, though formal negotiations have not yet begun.
Source: CNBC · 2026-09-16
Editor's Comment: Previously companies sought money; now money seeks companies. Capital's confidence in burn rates is more direct than any PPT.

2. Apple considers returning to server market, talks with Nvidia on networking

Summary: The Information reported that Apple is evaluating re-entering the server business and has held discussions with Nvidia regarding its networking technologies.
Source: The Information · 2026-09-16
Editor's Comment: Looking back twenty years after exiting servers, what Apple abandoned then happens to be the most valuable business today.

3. AI coding startup Factory triples valuation to $5 billion

Summary: The Wall Street Journal reported that AI coding company Factory's latest round valued it at over $5 billion, more than doubling from the previous round.
Source: WSJ · 2026-09-16
Editor's Comment: Coding agents are widely recognized as the "shovel sellers" track in this year's funding market. Behind the doubled valuation is every company equipping itself with AI programmers.

4. Google opens smart home to all AI agents

Summary: Google integrated smart home access via MCP-like protocols, allowing external AI tools like Claude to read and control users' connected devices and analyze household data.
Source: The Verge · 2026-09-17
Editor's Comment: Before handing your house keys to third-party agents, think about whose cleaning service you'd let into your bedroom.

5. GitHub's AI agent tricked into leaking private repos

Summary: Security firm Noma demonstrated an attack chain named GitLost, using crafted prompt content to induce GitHub's AI agent to leak user private repository information.
Source: Noma Security · 2026-09-16
Editor's Comment: Prompt injection has moved from papers to product lines. The attack surface of code hosting platforms has shifted from humans to agents.

6. AI agents get a "whistleblower hotline"

Summary: Two newly launched AI hotlines allow agents to directly report observed anomalous behaviors to security researchers. One was established by Ryan Greenblatt, a researcher at Redwood Research.
Source: TechCrunch · 2026-09-15
Editor's Comment: Having humans report machine colleagues is magical enough. Having machines report machines means the security industry has changed protagonists.

7. AI scams in dating apps are now mass-produced

Summary: A security researcher received a pop-up "confession" while analyzing the fraudulent dating app Dora. The Verge investigation revealed that AI-driven "pig butchering" dating apps have formed an industrial chain.
Source: The Verge · 2026-09-16
Editor's Comment: Scammers are the earliest group in the industry to achieve ROI with large models. The "user experience" iteration in this track is faster than legitimate products.

8. Perplexity builds storage engine CobbleDB with two engineers and AI agents

Summary: Perplexity published a blog post introducing CobbleDB, a storage engine developed by 2 engineers working with AI agents, revealing their internal development process.
Source: Perplexity · 2026-09-16
Editor's Comment: System software was considered the last bastion of AI programming. Now the bastion can be breached by teams of two.

9. Salesforce CEO denies "SaaS Apocalypse," Jensen Huang agrees

Summary: Marc Benioff called the claim that "AI will kill SaaS" nonsense. Jensen Huang expressed agreement, stating that AI will make existing software more valuable.
Source: TechRadar · 2026-09-16
Editor's Comment: When the subject of revolution personally debunks rumors, credibility takes a hit. But SaaS billing models are indeed being rewritten per-use by agents.

10. 404 Media: Agents are turning the internet into a server room where machines talk to each other

Summary: 404 Media reported that massive numbers of AI agents are beginning to bulk-produce content, comments, and traffic, rapidly diluting interaction data from human creators.
Source: 404 Media · 2026-09-16
Editor's Comment: When "user engagement" no longer distinguishes between humans and machines, the pricing unit of the entire advertising industry becomes invalid. That is the real bad news.

III. Humanoid Robots

1. Agility's new robot "squats to avoid" when encountering colleagues

Summary: Agility Robotics demonstrated the safety behavior logic of its new generation humanoid robots. When detecting a human colleague in a dangerous path, the robot actively stops and squats to avoid collision injuries.
Source: Ars Technica · 2026-09-16
Editor's Comment: The closer warehouse robots get to humans, the more safety protocols resemble insurance clauses. "Being able to squat" first convinces the employer's legal team, then the workshop floor.

2. Third World workers rent out bodies via forehead-mounted phones to feed humanoid robots

Summary: Data companies like Micro1 recruit workers in over fifty countries including Nigeria and India, buying first-person household video for about $15/hour as training material for companies like Tesla Optimus and Figure AI. Robot companies' annual spending on such data procurement has exceeded $100 million.
Source: Wired Investigation · NetEase Tech Translation · 2026-09-16
Editor's Comment: The data hunger of embodied intelligence penetrates down to the world's lowest-income populations. No one has quoted the moral bill for this industrial chain yet.

IV. Autonomous Driving

1. May Mobility goes public via $1.4 billion SPAC deal

Summary: Autonomous shuttle company May Mobility announced a merger with a special purpose acquisition company (SPAC), valuing the transaction at $1.4 billion, making it a listed company.
Source: TechCrunch · 2026-09-16
Editor's Comment: Companies unable to exit the primary market are queuing up through the SPAC back door. Going public here is a liquidity survival plan.

2. NHTSA gives Tesla deadline to explain Cybercab certification issues

Summary: The US National Highway Traffic Safety Administration required Tesla to answer a series of questions by September 30, explaining whether the steering-wheel-free, pedal-less Cybercab has completed proper self-certification.
Source: ITHome citing Reuters · 2026-09-16
Editor's Comment: Regulators aren't blocking new cars, only requiring documentation of "who is responsible for a car without brake pedals." This questionnaire is the first road test before mass production.

3. Former Waymo CFO jumps ship to Wayve

Summary: TechCrunch reported that Waymo's former Chief Financial Officer has joined UK autonomous driving startup Wayve.
Source: TechCrunch · 2026-09-16
Editor's Comment: When the person managing money at a top-tier company flows to a company that needs money, this signal indicates industry trends earlier than any technical evaluation.

4. Intel's multiple solutions exposed for edge AI agent autonomous driving

Summary: SDxCentral reported that Intel is embedding security capabilities into edge AI agents, optimizing ride-hailing fleet performance for autonomous driving scenarios.
Source: SDxCentral · 2026-09-16
Editor's Comment: Chipmakers are shifting from "selling compute" to "selling security layers." The second growth curve for edge computing is acting as bodyguards for agents.

V. World Models / Physical AI

1. MoE Capital long-form article dissects the "World Model" billion-dollar gamble

Summary: MoE Capital published a deep analysis tracing the convergence of world models from reinforcement learning and computer vision. It noted that over $10 billion has flooded into the field in the past 18 months, but robot AI maturity is far lower than what the funding scale suggests.
Source: MoE Capital · 2026-09
Editor's Comment: Fortunately, this article speaks the truth. Autonomous driving simulation is already profitable, while robot control is still early stage. One term cannot cover two levels of maturity.

2. Nvidia showcases WorldModelBench and other physical AI papers at NeurIPS 2026

Summary: Nvidia published its list of NeurIPS 2026 papers, including WorldModelBench, which judges video generation models directly on "whether they can serve as world models."
Source: NVIDIA · 2026-09-16
Editor's Comment: Generating something realistic and using it for decision-making are two different things. Evaluation benchmarks shifting to "usability" indicates this track is starting to hand in homework.

3. World Model Industry Map: Capital prices in advance, gap between valuation and revenue is biggest concern

Summary: Z Potentials industry report statistics show World Labs valued at ~$5.4 billion, Decart at ~$4 billion, and Yann LeCun's AMI Labs completing a $103 million seed round at a pre-money valuation of $3.5 billion. Total funding for pure world model startups in 2026 has exceeded $3 billion; most companies remain stuck at API calls and early pilots.
Source: Z Potentials · 2026-09-16
Editor's Comment: Money has priced them as "infrastructure," but revenue is still stuck at "project-based." This time lag is the bet for the next round of reshuffling.

Track Statistics: Large Models 8 items · AI Software 10 items · Humanoid Robots 2 items · Autonomous Driving 4 items · World Models/Physical AI 3 items, total 27 items.

Frontiers of the Physical World | WestTide

Tracks

I. Large Language Models (LLM / Foundation Model)

II. AI Software (AI Application / SaaS)

III. Humanoid Robots

IV. Autonomous Driving

V. World Models / Physical AI

2 replies