Today's roundup, collected through the evening of November 6 (Japan time), paired robots edging toward practical use with a market growing cold. In San Francisco, a humanoid robot cleaning service has launched, and a first-hand account of it is making the rounds. At the same time, capital markets and corporate strategy made plenty of headlines, from reports of Anthropic's IPO troubles to Tesla discontinuing the Cybertruck.

A Humanoid Cleaning Service Comes to San Francisco — Teleoperated, but "a Taste of the Future"

A reporter at MIT Technology Review tried the humanoid cleaning service that Tau Robotics recently launched in San Francisco, booking a slot immediately after seeing a post on X.

For now, the robot isn't fully autonomous — a human operator teleoperates it via cameras mounted on its head and wrists. The article describes the setup as "closer to a radio-controlled car than a Waymo," and for tasks like making the bed or folding laundry, assistance from a human "motion guide" is apparently also required.

What deserves attention, though, is the data collection. Everything the robot does while cleaning is recorded as Tau training data and used to train more autonomous next-generation models. Every wiped-down counter and folded T-shirt becomes teaching material that makes the next model smarter. Because that data is so valuable, pricing is held at a level you'd never guess was cutting-edge: $30 for one hour at the time of the review, now $60 for two robots once a week, with one-off jobs at $100.

The robot's look is unique too. It stands roughly 140 cm tall. Its 3D-printed head sports an antenna for communicating with headquarters, and in place of human-like hands it has claws. Humanlike hands weren't durable enough, and the claws actually turned out to be better for most cleaning tasks, the article says. To keep from scratching the flooring, the robot wore grippy socks. The verdict from the experience: thorough but slow — on cost-performance alone, you'd be better off cleaning yourself. Still, as an hour spent getting a taste of a Jetsons-style future, it seems the author found it worthwhile.

Anthropic's $1.5 Trillion IPO Already Hitting Turbulence

According to The Decoder, Anthropic's IPO plans are running into trouble. Investors are balking at the $1.5 trillion valuation, and a broad sell-off in tech stocks is adding to the pressure.

The past few days' plunge in AI stocks (which we've covered here day after day) is chilling the environment around large IPOs. As we noted recently, OpenAI's Greg Brockman said the company will be "in a position to IPO" in 2027 — but a giant valuation at a moment when the market has cooled means investors will scrutinize the deal all the more harshly. For now, details such as the size and timing of the offering are unknown; whether talks continue or the valuation gets revised bears close watching.

Tesla Kills the Cybertruck — Musk Pivots to Robotics and AI

The Financial Times reports that Tesla will end production of the Cybertruck electric pickup and that Elon Musk is shifting the company's strategic focus to robotics and AI (the story took off via Hacker News).

For Tesla, which has been developing robotaxis and its Optimus humanoid, it has long been clear that robotics and AI are meant to be the next pillar. With a commercial service of robots actually working inside homes now up and running, a picture is emerging in which incumbents and startups alike are chasing the "robot in your home."

Cohere: "Produce More Gas, for AI's Sake" — Power Demand Ignites a Debate

According to a November 4 report from DeSmog, generative AI company Cohere has argued that producing more gas is necessary to run more AI. With the data center boom pushing electricity demand up against generation capacity, the AI industry is now the one calling for expanded fossil fuel production.

Around AI's power problem, investment is in fact still growing — Google, for example, is planning a roughly $17 billion data center in Wyoming. With demand outstripping supply, it's a natural drift for natural gas to join nuclear and renewables in the discussion, but how this squares with climate policy looks likely to become a future point of contention.

Security Capabilities Disappear from China's Open-Weight Models

As reported in Reddit's LocalLLaMA community, Kimi K3.1 and Tencent's Hy4 have removed cybersecurity, reverse-engineering, and vulnerability-patching capabilities from their open-source models. Tencent's research page also reportedly mentions its approach to "potential attack capabilities."

Hy4, which we introduced here the other day, is a strong open model with image understanding and computer use that staked a claim as "No. 2 among the Chinese contenders." But in the open release, capabilities useful for security research may have been deliberately limited. The poster calls it "a bad pattern being repeated across China's latest open-weight models." The fallout for people using local LLMs for security research — and just how much of this is intentional restriction — is something we'll keep an eye on.

Inkling 2 Small's Training Details — 68 Hours and 12T Tokens to Match or Beat Chinese Models

Latent Space's AI News has the training details on Thinking Machines' Inkling 2 Small. The model, with 17B active and 594B total parameters, was pretrained from scratch in NVFP4 on under 12T tokens in 68 hours (less than three days) and finished with a single joint RL run. It reportedly goes toe-to-toe with — and often beats — Chinese models of the same size, and its vision encoder is also rated as extremely high quality.

Citing results like Tinker's first anniversary and Interaction Models, the newsletter describes Thinking Machines as "a neolab delivering strong results at the American frontier." It may signal the arrival of an era in which a small, elite lab can close in on the giants with efficiency-first training.

Other Moves

JFrog is forecasting its strongest growth in four years (Bloomberg). Security demand is climbing as AI hands hackers advanced techniques, and the CEO said there is "panic about what happens if we don't protect the software supply chain."

Anthropic has published "AI for scientific discovery," a report rounding up how researchers are using Claude for science.

The AP reports that Anthropic is rapidly accelerating AI development even as calls to slow down grow louder.

A post on LocalLLaMA claims "Qwen 4 has been stealth-announced," based on posts from Alibaba's official Qwen account. Information is still limited.