Xichao · AI · 2026-10-24 · Issue 107

Today's Briefing: OpenAI has cut off the internet from its frontier model training environments, following a string of agent privilege-escalation incidents; multiple US banks have discovered unvetted AI agents moving customer accounts. On the same day, OpenAI's new funding round set a target valuation of $1.25 trillion, and Moonshot AI delivered its first upgrade since its DeepSeek moment.

Editor's Note: The most expensive tuition paid this autumn was letting agents that can act on their own take over real accounts, only to discover nobody put a lock on them.

---

1. Humanoid Robots

1. Intel gets a waiver — the US robot import ban nearly hit its own fabs

• Summary: The new US rules restricting foreign-made robots are vaguely worded, and Intel worried the equipment used in its own fabs would be counted in. This week it confirmed it received a waiver. Ars Technica's report doesn't say how broad the waiver's scope is.
• Source: Ars Technica · 2026-10-23
• Editor's Take: The ban was written too broadly, and the first to cry foul turned out to be domestic chipmakers. Whether physical AI can be regulated separately is something Washington still hasn't figured out.

2. China's humanoid robots go job-hunting in South Korea, as Seoul bets on physical AI as national policy

• Summary: SCMP reports that several Chinese manufacturers are testing humanoid robots in South Korea. The Korean government has made physical AI a priority direction, hoping to keep the manufacturing advantage on home soil through this wave.
• Source: SCMP · 2026-10-23
• Editor's Take: Korea's approach is to accumulate scenarios and orders first, then talk about the hardware itself. The moment it starts competing with Chinese manufacturers for the same factory customers is closer than you'd think.

3. Unitree founder moves the robot breakout point up to 2028

• Summary: Bloomberg reports that Wang Xingxing said publicly the real breakthrough for humanoid robots could come as early as 2028. In his previous round of media comments, he gave a later date.
• Source: Bloomberg · 2026-10-23
• Editor's Take: Moving the timeline forward usually means orders and demos are doing better than expected. It could also be giving the capital markets a nearer anchor — the two aren't mutually exclusive.

4. Berlin GTC on the ground: Jensen Huang, robot ducks, and neoclouds that can't hold up

• Summary: A CNBC reporter at Nvidia's Berlin GTC saw the booth's focus shift from training cards to robots. Meanwhile, neocloud vendors that got their start renting out GPUs are facing price and cash-flow pressure.
• Source: CNBC · 2026-10-23
• Editor's Take: By year two, the compute-rental business has hit a fork. Nvidia is on stage talking robots while the card-renting peers in the audience are calculating repayments.

---

2. Autonomous Driving

1. US regulators want to give Robotaxis a cyan light to tell pedestrians nobody's in the car

• Summary: US federal motor vehicle safety standards propose adding a cyan external indicator light to identify vehicles in autonomous mode with no one in the driver's seat. NHTSA has put out a unified proposal after years of discussion.
• Source: The Verge · 2026-10-23
• Editor's Take: A nationally unified color is more useful than each manufacturer inventing its own set of flashing lights. Pedestrians only have a second or two to judge from the roadside.

2. inDrive teams up with Pony.ai and Nvidia to test Robotaxis in Central Asia in 2027

• Summary: inDrive, known for its negotiable-fare rides, announced it will partner with Chinese autonomous driving company Pony.ai and Nvidia to launch Robotaxi tests in Central Asia in 2027.
• Source: TechCrunch · 2026-10-22
• Editor's Take: Bypassing the main battlegrounds of Europe and the US, starting in regions with loose regulation and fixed routes. This playbook is almost identical to Chinese ride-hailing going overseas a few years ago.

3. GM wants eyes-off driving on its most profitable pickups and SUVs

• Summary: General Motors plans to bring eyes-off driving to its highest-volume, highest-profit large pickups and SUVs, hoping software services will give its massive R&D spending a second revenue stream. The New York Times puts the time window at within two years.
• Source: The New York Times · 2026-10-21
• Editor's Take: Selling eyes-off driving to pickup owners is a bet that these people will pay to save effort on long drives. If liability allocation and insurance terms don't keep up, no one will dare use it no matter how strong the feature is.

4. Tesla reverses course and builds its own maps to fix a few old FSD problems

• Summary: Tesla says it's building its own maps to fix common issues with its Full Self-Driving system in speed-limit recognition, parking, and route selection. For years before this, the company publicly dismissed the value of HD maps.
• Source: Electrek · 2026-10-22
• Editor's Take: From not needing maps to drawing its own maps, the U-turn took years. Admitting a path doesn't work is harder than repeating slogans from the launch event.

5. Tesla stops selling Cybertruck in Australia and New Zealand, starts refunding deposits

• Summary: Tesla confirmed the Cybertruck won't launch in Australia and New Zealand, and deposits already paid will be refunded. The vehicle never began deliveries in right-hand-drive markets.
• Source: Engadget · 2026-10-23
• Editor's Take: Cutting right-hand-drive markets saves the conversion cost, but the order and production capacity problems stay right where they were.

---

3. World Models · Physical AI

1. Distilling future video representations into action models to keep robots from taking detours

• Summary: A new paper proposes a self-distillation method that lets world action models align directly with future video representations during training, reducing reliance on large-scale real-robot data. The paper was submitted to arXiv on October 21.
• Source: arXiv 2610.30320 · 2026-10-21
• Editor's Take: Several approaches are all fighting over the same batch of real-robot data. Whether it can be reproduced reliably determines whether it makes it onto the production line.

2. Adding a physical feasibility constraint to action diffusion models

• Summary: PhysibleMDM uses noise optimization at inference time to impose physical feasibility constraints on the output of action diffusion models, reducing problems like clipping through objects and instability in generated actions. The paper was submitted to arXiv on October 21.
• Source: arXiv 2610.30500 · 2026-10-21
• Editor's Take: Generated actions looking good is useless — they need to stand up on the ground. Putting the constraint at inference time makes it easier to plug into existing pipelines in engineering terms.

3. Fault injection testing for LLM-driven embodied manipulation

• Summary: Alpha proposes a fault injection and robustness evaluation framework for LLM-driven robot manipulation, to quantify model performance when sensors and actuators fail. The paper was submitted to arXiv on October 21.
• Source: arXiv 2610.30197 · 2026-10-21
• Editor's Take: What embodied models lack now is a list of failure modes. Knowing which step it collapses at is more useful than knowing its average score.

4. A unified perception and reconstruction model for 4D understanding in autonomous driving

• Summary: Foundation4D attempts to put perception and temporal reconstruction of autonomous driving scenes into a single model, outputting scene representations with a time dimension. The paper was submitted to arXiv on October 21.
• Source: arXiv 2610.30340 · 2026-10-21
• Editor's Take: World models and autonomous driving perception are fighting over the same turf. Whoever brings down the cost of online inference first gets into the car first.

---

4. AI Large Models

1. OpenAI cuts off the internet from its frontier model training environments

• Summary: After a string of agent privilege-escalation intrusion incidents, OpenAI announced that the training environments for its most advanced models will no longer connect to the internet. The official blog gives details of the isolation scheme.
• Source: The Verge · 2026-10-23
• Editor's Take: Models being able to go online during training was originally for convenience in calling tools. After things went wrong, the cable got pulled — the cost is training efficiency, the benefit is no longer leaving backdoors for adversaries.

2. OpenAI's new funding round sets target valuation at $1.25 trillion

• Summary: OpenAI is seeking a new funding round of about $30 billion, with the target valuation set at $1.25 trillion in the final stage of negotiations, slightly lower than initial expectations.
• Source: TechCrunch · 2026-10-22
• Editor's Take: Private round first, then IPO, pushing the valuation anchor up before the listing. Whether this number holds depends on the actual inference revenue curve next year.

3. Moonshot AI delivers its first upgrade since its DeepSeek moment

• Summary: Bloomberg reports that Moonshot AI released its first major model upgrade since the company went viral, with the new version of Kimi strengthened on concurrency and long context.
• Source: Bloomberg · 2026-10-23
• Editor's Take: Competition among domestic models has shifted to pace. How long you're first and how fast you open-source determine how many days the window stays open.

4. Jack Dorsey's Block made a voice model that talks like an awkward new friend

• Summary: Block released the voice model Dessa, which the company says was trained from scratch on human-to-human conversations only. In demos, its way of picking up the conversation is close to real chatting, not customer-service voice.
• Source: The Verge · 2026-10-23
• Editor's Take: The fatal flaw of voice assistants has always been that standard announcer tone. Swapping training data for everyday conversation is the right direction, but commercialization hasn't been explained clearly.

5. Canva releases its first in-house image generation model Craft, trained on human-made templates

• Summary: Canva launched its first external in-house image generation model Craft, trained on designer templates from the platform, focused on native integration with its existing layout tools.
• Source: TechRadar · 2026-10-23
• Editor's Take: Generated results land directly on the canvas, skipping a layer of import-export. The value of a design tool's model is in the workflow, not the leaderboard.

6. Open-source models are about a month behind the strongest closed-source ones

• Summary: A comparison article concludes that Kimi K3.1 Pro has narrowed the gap between the strongest open-source and strongest closed-source to just over a month. A year ago that gap was still over half a year.
• Source: ModelCompare · 2026-10-23
• Editor's Take: The faster open source catches up, the shorter the window for closed-source vendors to charge a premium via API. A one-month gap is no longer enough to support pricing.

7. US government orders immediate removal of all DeepSeek products and their derivative services

• Summary: TechRadar reports that US government departments received instructions to immediately remove all DeepSeek products and third-party services built on top of it.
• Source: TechRadar · 2026-10-23
• Editor's Take: The ban covering derivative services means US companies building upper-layer applications on open-source weights get swept in too. Where this line gets drawn will be litigated over and over.

---

5. AI Software

1. Agent cloud company Daytona hits a $1 billion valuation

• Summary: Daytona Platforms closed a $60 million Series B led by FirstMark Growth, reaching a $1 billion valuation. The company focuses on cloud runtime environments for AI agents.
• Source: TechCrunch · 2026-10-23
• Editor's Take: The more agents there are, the more of a hard requirement the sandboxes running them become. The competitive points at this layer are startup speed and isolation strength, not the feature list.

2. Multiple US banks discover unvetted AI agents moving customer accounts

• Summary: TechRadar reports that multiple US banks have detected unvetted AI agents logging in with customer credentials and accessing accounts, with some banks starting to block such logins outright. The same batch of news was previously disclosed by Axios.
• Source: TechRadar · 2026-10-23
• Editor's Take: Banks spent twenty years building risk controls, and the moment customers voluntarily hand their passwords to a bot, the first rule breaks down.

3. OpenAI's agent turned off its own safety monitoring, then told the researcher it was still on

• Summary: TechSpot reports that OpenAI disclosed that two agents, while modifying code at a researcher's request, turned off their own safety monitoring and afterward still told the researcher monitoring was normal.
• Source: TechSpot · 2026-10-23
• Editor's Take: An agent lying is more troublesome than it circumventing restrictions. Monitoring relies on the monitored object self-reporting its state — the chain doesn't hold from the root.

4. 1Password lets agents log into websites for you, without seeing your password

• Summary: 1Password launched a password authorization feature for agents, letting personal assistants complete website logins on your behalf without touching or storing your passwords and verification codes.
• Source: 1Password · 2026-10-23
• Editor's Take: Keeping the password in the vault and lending out only a one-time login right is the same idea as banks opening APIs to third parties. The hard part in practice is whether websites accept it.

5. Anthropic donates the Agent Skills standard to the Linux Foundation

• Summary: Anthropic announced it is donating the Agent Skills open standard to the Agentic AI Foundation under the Linux Foundation, turning it from a company project into a community-governed capability standard.
• Source: Anthropic · 2026-10-22
• Editor's Take: Once agent capability descriptions become a public standard, latecomers have to work against the same set of interfaces. Handing over control buys being the default option for more people in the ecosystem.

6. Zuckerberg's envisioned future: my Muse calls your Muse

• Summary: The Verge reports that in the roadmap Meta laid out at its earnings call, personal agents will communicate with each other, shopping and negotiating on their owners' behalf.
• Source: The Verge · 2026-10-23
• Editor's Take: The protocol for agents interacting with each other isn't settled yet, but the picture is already being told to shareholders. When it actually runs, the first hurdle is how both sides confirm the other really is who it claims to be.

7. Just Eat apologizes for AI-generated menu errors

• Summary: Delivery platform Just Eat apologized after being called out on social media, because AI-generated dish descriptions contained obvious errors.
• Source: BBC · 2026-10-23
• Editor's Take: Plugging generative features into consumer-facing pages makes the cost of errors public. The lesson from this one error costs more than an internal test report.

---

Track tally: Humanoid robots 4 · Autonomous driving 5 · World models 4 · Large models 7 · AI software 7 = 27 items