WestTide · AI · 2026-11-06 · Issue 120

Today's Briefing: Agility Robotics lists on Nasdaq, giving the US its first pure-play humanoid robot public company; OpenAI's cybersecurity model combed through the Linux kernel, confirming and patching nearly 1000 issues; Waymo confirms it will end its exclusive partnership with Uber in Austin and Atlanta starting January 2028.

Editor's Note: Today's two headlines are worth reading side by side. One robotics company now has to answer to quarterly earnings; one model company has started selling the ability to "catch bugs." Capital and security are pricing this industry at the same time, and neither is waiting for the other.

---

1. Humanoid Robots

1. Agility lands on Nasdaq, giving American humanoid robots their first pure-play ticker

• Summary: Bloomberg reports that Nvidia-backed Agility Robotics listed on Nasdaq this week, becoming the first US public company whose main business is humanoid robots. TechCrunch's piece the same day put the question in the headline: are you buying? Orders, gross margins, delivery cadence — numbers that used to appear only in roadshow materials — now have to be disclosed quarterly starting today.
• Sources: Bloomberg · TechCrunch · 2026-11-05
• Editor's Take: Going public solves financing, not customers. Most humanoid orders right now come from pilots, and no one has publicly disclosed what share of pilots convert to mass production.

2. Xiaomi's humanoid robot works an 8-hour shift at its own car factory

• Summary: Xiaomi released in-factory footage of a humanoid robot sorting center-console side panels on a full vehicle production line, working continuously for 8 hours. The company says the robot completed everything autonomously and even handled an incoming-material anomaly mid-shift.
• Sources: Electrek · 2026-11-04
• Editor's Take: Factories do their math by shift — it only makes sense as a replacement if it can cover a full shift. The video gives no cycle time or yield rate, and those two are what decide whether it's still on that line next year.

3. Serve Robotics wants to sell delivery-robot operations as a cloud service

• Summary: Serve, which makes sidewalk delivery robots, announced a new business packaging its autonomous driving stack, fleet dispatch, and operations for external customers on a usage-based fee. The goal is to turn itself from a hardware seller into an infrastructure provider.
• Sources: TechCrunch · 2026-11-05
• Editor's Take: The most expensive part of delivery robots is operations per kilometer, not the algorithm. Renting out operations is an admission that the ceiling for a self-operated fleet is lower than expected.

4. China blocks SoftBank's $5.4B acquisition of ABB's robotics business

• Summary: Both The Information and Bloomberg report that Chinese regulators did not clear SoftBank's $5.4 billion acquisition of ABB's industrial robotics division. ABB robots are installed on a large number of production lines in China, and the deal is stuck at the approval stage.
• Sources: The Information · Bloomberg · 2026-11-05
• Editor's Take: Industrial robots are the infrastructure of manufacturing. Assets like these rarely change hands on purely commercial logic.

5. Toyota sets up a robotics division, consolidating scattered teams

• Summary: Toyota established a dedicated robotics division, consolidating R&D resources previously spread across vehicles, components, and its research institute. Bloomberg says the move tracks with other automakers' robotics pushes over the past two years.
• Sources: Bloomberg · 2026-11-05
• Editor's Take: Automakers' edge in robotics is supply chain and mass-production know-how; their weakness is software. Whether consolidating the org can fix the latter will be clear within a year.

6. Japan's Taisei: AI-operated tunnel boring machine beats veteran operators

• Summary: Nikkei reports that Taisei used AI to autonomously operate a giant shield tunneling machine in a demonstration, outperforming skilled operators. Shield steering has long relied on a feel for the geology — expertise that used to live only in people's hands.
• Sources: Nikkei Asia · 2026-11-05
• Editor's Take: Stable conditions, dense data, quantifiable cost of error — this is a rare virtuous loop for physical AI. It'll see returns before humanoid robots do.

---

2. Autonomous Driving

1. Waymo confirms split with Uber in January 2028

• Summary: Waymo co-CEO Tekedra Mawakana told Bloomberg that the company will end its exclusive partnership with Uber in Austin and Atlanta starting January 2028, after which Waymo vehicles can join other platforms.
• Sources: TechCrunch · 2026-11-04
• Editor's Take: Exclusivity traded for scale, and Waymo chose the latter. Before 2028 it has to prove a single car can make money; otherwise more channels just spread the losses around.

2. Tesla sold tens of thousands of FSD packages in Korea, and owners still can't use them

• Summary: A Korean consumer advocacy group says Tesla has sold FSD option packages to tens of thousands of local owners even though the feature isn't available there, and is demanding refunds or a clear timeline.
• Sources: InsideEVs · 2026-11-05
• Editor's Take: Prepaid software becomes a liability in markets where regulation lands slowly. Tesla has faced similar complaints in several markets.

3. A startup turns Teslas into 11-seat minibuses and pickups

• Summary: Hoffman-Badger does full-vehicle conversions, stretching Model 3 and Model Y bodies into 11-seat minibuses, pickups, and other forms while keeping the original powertrain and electronics architecture.
• Sources: InsideEVs · 2026-11-05
• Editor's Take: The conversion business feeds on platform commonality — and it also exposes how slow the OEM has been on niche use cases.

4. Language world model takes over driving decisions; new paper emphasizes interpretability

• Summary: A new paper uses a language model as a driving world model, first reasoning in natural language about other vehicles' intentions and risks, then converting that into control commands. The authors say it's easier to trace than end-to-end in long-tail scenarios.
• Sources: arXiv · 2026-11-03
• Editor's Take: Interpretability is the knock on regulators' door; the price is inference latency. Whether this road gets into cars depends on how fast the model runs.

---

3. World Models · Physical AI

1. Nvidia open-sources Cosmos 3, world models start reaching developers

• Summary: Nvidia released the Cosmos 3 family of open models on Hugging Face, focused on interactive world simulation: given an action, it predicts how a scene evolves, aimed at training and evaluation for robotics and autonomous driving.
• Sources: Hugging Face · 2026-11-06
• Editor's Take: Open-sourcing world models pushes the marginal cost of synthetic data close to zero. The gap between simulation and real machines will get partly filled in by a lot of trial and error.

2. Surgical world model SurgTransfer: transferring manipulation skills via latent actions

• Summary: The paper proposes multi-step latent action modeling, learning action representations from surgical video and transferring them across instruments and viewpoints for prediction and planning in a surgical world model.
• Sources: arXiv · 2026-11-03
• Editor's Take: Medicine is the lowest-tolerance, most expensive-data scenario in physical AI. The selling point of latent actions is less labeling.

3. OdorSpaces: letting robots navigate 3D environments by smell

• Summary: The paper builds an embodied navigation framework combining vision and olfaction, using both images and odor concentration fields to locate targets in 3D scenes — an exploration in multimodal embodied perception.
• Sources: arXiv · 2026-11-03
• Editor's Take: The olfactory channel has long been ignored; industrial inspection and search-and-rescue are the few scenarios where it holds up.

---

4. AI Large Models

1. OpenAI releases GPT-6 Cyber, and combs through the Linux kernel along the way

• Summary: OpenAI launched GPT-6 Cyber for cybersecurity, claiming it finds more vulnerabilities and oversteps less. In the accompanying report, the automated system flagged 2308 suspects in the Linux kernel, dynamically reproduced 1724, and confirmed and patched nearly 1000 issues — a third of which had already been discussed publicly.
• Sources: The Verge · TechRadar · 2026-11-05
• Editor's Take: Vulnerability hunting is shifting from people-plus-tools to models-plus-pipelines, and security teams' headcount structure will change first. Patch generation is the real barrier.

2. Claude connects to bank accounts and starts managing your money

• Summary: Anthropic launched Personal Finance, connecting Claude to users' bank and credit card accounts to read transactions, categorize spending, and suggest budgets. It's in small-scale testing in the US.
• Sources: The Verge · TechCrunch · 2026-11-05
• Editor's Take: Financial data is the ceiling of trust. Taking this step moves Claude from assistant to agent, and who's responsible when it errs has to be rewritten.

3. Gemini Omni 1.2 Flash bets on cheap and long clips

• Summary: Google released Gemini Omni 1.2 Flash, officially billed as faster and cheaper, supporting longer video clip generation and editing, billed per usage for developers.
• Sources: Google · 2026-11-06
• Editor's Take: The video generation price war has moved from seconds to cost per frame, and small models like Flash are the main force.

4. Moonshot AI goes via Europe to find Blackwell

• Summary: CNBC, citing sources, says Moonshot AI — which the White House alleged remotely accessed high-end Nvidia chips in East Asia — is seeking Blackwell compute in Europe.
• Sources: CNBC · 2026-11-05
• Editor's Take: One of a model company's core competencies now is whether it can reliably get cards. That's not on the technical roadmap, but it decides whether the roadmap gets finished.

5. Apple puts a budget on employees' AI usage

• Summary: Bloomberg reports Apple has started capping daily AI call quotas for engineering teams, folding token consumption into cost management. TechRadar says new CEO John Ternus has tightened such spending since taking over.
• Sources: TechCrunch · TechRadar · 2026-11-04
• Editor's Take: AI has moved from "use it" to "account for it." What gets measured next is output per dollar.

6. Ads in chatbots: the camps are splitting

• Summary: Wired surveyed the various forms of chatbot advertising. Anthropic insists Claude will carry no ads, and its IPO filing has no advertising revenue line; another group of vendors has already sold off recommendation slots.
• Sources: Wired · 2026-11-06
• Editor's Take: No ads is a differentiator, and also a cost line. The moment subscription growth slows, that promise gets repriced.

7. Open models are growing — who's making the money

• Summary: TechCrunch talked with Hugging Face's policy lead about the costs, risks, and business paths of open models. One conclusion: enterprise spending is shifting from API bills to inference infrastructure.
• Sources: TechCrunch · 2026-11-05
• Editor's Take: Open source pushes model prices down and raises the ops bar. The money moves from selling models to selling what makes them run.

8. OpenAI president: could be IPO-ready in 2027

• Summary: Greg Brockman said the company could be ready for an IPO in 2027, with enterprise revenue growth as the main support.
• Sources: TechCrunch · 2026-11-05
• Editor's Take: Anthropic files, OpenAI talks — both turned the conversation toward public markets in the same week. Valuation needs a revenue curve, and the revenue curve needs enterprise renewals.

---

5. AI Software

1. Dropbox buys Tigris Data for $350M

• Summary: Dropbox acquired cloud storage company Tigris Data for $350 million, folding object storage into its own AI infrastructure — its largest technology acquisition in recent years.
• Sources: TechCrunch · 2026-11-05
• Editor's Take: A cloud-drive company buying the storage layer is about keeping AI feature margins in-house.

2. Instagram brings Edits to Mac, and you can plug in your own agent

• Summary: Edits launched a macOS desktop version where users can hook their chosen AI agent into the timeline to assemble videos; Meta uses MCP to connect external tools. The mobile app already has over 50 million users.
• Sources: The Verge · 2026-11-05
• Editor's Take: An editing tool adopting MCP shows the fight over agent entry points has moved from chat boxes into productivity software.

3. Etsy requires at least one real photo per listing

• Summary: Etsy updated its rules: listings must include at least one photo of the actual item, and AI-generated images and descriptions are restricted. The platform says this is so buyers can judge the real thing.
• Sources: TechRadar · 2026-11-06
• Editor's Take: Platforms draw the line on AI content using verifiable photographic evidence. Enforcing this rule won't be cheap.

4. GM recalls nearly 70,000 EVs over brake software defect

• Summary: GM recalled more than 67,000 Cadillac and Chevrolet EVs over a brake software defect that could cause brief braking failure. In the same period, Hyundai recalled nearly 30,000 Kona Electrics over an airbag software issue.
• Sources: Electrek · TechRadar · 2026-11-04
• Editor's Take: The flip side of software-defined cars is that software defects are also priced per vehicle. OTA can fix it, but recall notices still go out.

5. Uber: almost no one books rides with AI agents

• Summary: On its Q3 earnings call, Uber said the share of orders placed through personal AI agents is negligible, and the company is still testing agent integration.
• Sources: TechCrunch · 2026-11-04
• Editor's Take: The "agents replace apps" story has been told for two years; without solving payments and identity, agent booking stays in demo land.

6. METR assesses Anthropic's internal AI R&D acceleration, some content redacted

• Summary: METR published an independent assessment of Anthropic's AI R&D acceleration as of October 2026. At the company's request, non-public information was redacted; the public portion covers methodology and the conclusion framework.
• Sources: METR · 2026-11-05
• Editor's Take: If only the methodology can be published, these assessments already touch trade secrets. Third-party auditing of AI R&D speed still lacks a disclosable standard.

7. Aina smart ring: a remote control for your agent

• Summary: Aina launched a smart ring with a screen, joystick, and microphone, used to approve or reject agent requests. Wired calls it the remote control of the agent era.
• Sources: Wired · 2026-11-05
• Editor's Take: Agents want permissions, and people need a physical switch. Whether this category works depends on whether it's faster than pulling out your phone.

---

Track tally: Humanoid robots 6 · Autonomous driving 4 · World models & physical AI 3 · AI large models 8 · AI software 7 — 28 items total.

Physical Frontier | WestTide