Physix FrontierWestTide · AI · 2026-10-25 · Issue 108
Today's Briefing: OpenAI cuts off the external network for frontier model training environments, on the grounds that model behavior is still unpredictable; NHTSA proposes hanging a unified cyan-green light on autonomous vehicles so passersby can tell at a glance there's nobody in the car; Block releases Dessa, a voice model trained from scratch only on human conversations. On the capital side, OpenAI's new funding round targets a valuation of $1.25 trillion, Anthropic's IPO is imminent, and early investors' paper numbers have already ticked up a round.
Editor's Note: The main thread overseas this week is one thing — putting locks on agents that can act on their own. The locks aren't installed yet, but the bill has already arrived.
1. Humanoid Robots
1. Unitree talks up robots that "self-evolve," Tesla raises Optimus production target to 20,000 units per week
- Summary: Unitree says "self-evolving robots" are the next leap for humanoids. Meanwhile Tesla has raised Optimus's production target to 20,000 units per week, and the model is already in production at the Fremont factory.
- Source: TechRadar · 2026-10-24
- Editor's Take: Both are fighting for the mindshare around the same thing — whether per-unit cost can be pushed below the line where it makes sense to put them to work in factories. Whoever draws that line first is the one whose production numbers actually mean something.
2. Jensen Huang plays with robot ducks at GTC Berlin, 20,000 units sold out
- Summary: At Nvidia GTC Berlin, Jensen Huang interacted with a bunch of waddling Microduck robots during his keynote. After the video circulated, about 20,000 units of these little robots sold.
- Source: TechCrunch · 2026-10-23
- Editor's Take: A toy selling out doesn't say anything about industry progress, but it does shove the term "physical AI" into ordinary people's shopping carts. That's exactly the exposure Nvidia wants.
3. US robot import ban wording is vague, Intel gets a waiver first
- Summary: The boundaries of the US's new rules restricting foreign-made robots are unclear. Intel worried fab equipment might be counted in, and confirmed this week it got a waiver. Ars Technica did not say how broad the waiver's scope is.
- Source: Ars Technica · 2026-10-23
- Editor's Take: The broader the rules are written, the more likely the first to cry foul are domestic chipmakers. Whether physical AI should be regulated separately is a question Washington hasn't answered yet.
2. Autonomous Driving
1. NHTSA wants to give autonomous vehicles cyan-green lights
- Summary: US NHTSA has proposed introducing a unified cyan-green external marker light for autonomous vehicles, to signal that the vehicle is in autonomous mode and no one is in the driver's seat.
- Source: Electrek · 2026-10-24
- Editor's Take: One nationwide color beats every company inventing its own flashing lights. Pedestrians only have a second or two to judge at the roadside.
2. Chinese EVs enter Western markets through the back door, selling 40% below MSRP
- Summary: Electrek reports that Chinese EVs are entering Western markets through more discreet channels, with transaction prices about 40% below manufacturer suggested retail price. Rising energy prices continue to push up the cost of new cars from Western domestic brands.
- Source: Electrek · 2026-10-24
- Editor's Take: Tariffs block the sticker price, not the channels. With the price gap this wide, dealers will find their own routes.
3. Autonomous driving progress on Tesla's earnings call
- Summary: Electrek's weekly podcast went through Tesla's Q3 earnings call, covering the timeline for autonomous driving features, and also touched on updates to the BYD Seagull and Mustang Mach-E.
- Source: Electrek · 2026-10-23
- Editor's Take: When they talk autonomous driving on an earnings call, investors are listening for revenue recognition criteria, not feature demos.
3. World Models · Physical AI
1. Rivian turns 400,000 video clips a day into 3D simulation, feeding autonomous driving with synthetic worlds
- Summary: Rivian acquired a synthetic data startup, converting the 400,000 clips its fleet uploads daily into 3D simulated scenes for training autonomous driving systems.
- Source: Electrek · 2026-10-23
- Editor's Take: Real-car data is expensive because it's scarce; simulation is expensive because it's cheap. Moving long-tail scenarios into 3D environments is the most direct lever for cutting training costs.
2. Open-source parametric CAD tool Loböle lets AI agents drive modeling directly
- Summary: A developer released the open-source parametric CAD project Loböle, built on OpenCASCADE for B-Rep modeling, with an interface
facing AI agents, aiming to let agents directly generate and modify mechanical structures.
- Source: Hacker News · 2026-10-24
- Editor's Take: Half of physical AI's bottleneck is in the models, the other half is whether the design toolchain leaves an interface for agents. Projects like this are still small, but the direction is worth watching.
4. AI Large Models
1. OpenAI announces frontier model training goes fully offline
- Summary: OpenAI says it will switch to training models on cached web content and simulated internet environments, no longer letting models connect directly to the real network, on the grounds that model behavior remains hard to predict. The company had just disclosed an agent privilege-escalation incident previously.
- Source: The Verge · 2026-10-24
- Editor's Take: Isolating the training environment is an admission that network access itself is an attack surface. This step should have been taken long ago; the cost is that models become duller in their perception of the current world.
2. UK invests $134 million to build an AI biosecurity lab
- Summary: The UK government announced an investment of about $134 million to build an AI biosecurity lab, with OpenAI, Anthropic, and Google all expressing support.
- Source: Engadget · 2026-10-24
- Editor's Take: Biosecurity is one of the few AI topics countries are still willing to fund together. It's not a lot of money, but the signal value outweighs the amount.
3. Block releases voice model Dessa, trained from scratch only on human conversations
- Summary: Jack Dorsey's Block released the voice model Dessa, which the company says was trained entirely from scratch on human conversation data, can run locally, and whose expression and pauses sound close to a real person.
- Source: The Verge · 2026-10-23
- Editor's Take: Local running plus real human conversation corpora — the selling points are privacy and naturalness. Competition in voice models is shifting from "does it sound real" to "does it run on your own device."
4. ChatGPT launches "locked chats," privacy terms leave a back door
- Summary: OpenAI rolled out a PIN-protected "locked chats" feature for ChatGPT. After reading through the fine print, TechRadar pointed out there are still several exception clauses beyond the encryption.
- Source: TechRadar · 2026-10-24
- Editor's Take: The credibility of a privacy feature depends on how long the exception list is. Marketing copy talks about encryption; what users should read are those provisos.
5. Study: when forced to pick sides, Grok picks Republicans, OpenAI picks Democrats
- Summary: A new study found that most chatbots won't proactively recommend who to vote for, but when forced into a binary choice, Grok's model leans Republican and OpenAI's model leans Democratic.
- Source: TechCrunch · 2026-10-24
- Editor's Take: The political undertone of training data can't be hidden when forced to take a stance. The value of studies like this is that they put the word "neutral" to the test.
6. Study: AI models think humans are a bunch of amoral sheep
- Summary: The Register reports that researchers found several mainstream models default to negative descriptions of humans, tending to portray crowds as a herd lacking moral judgment.
- Source: The Register · 2026-10-23
- Editor's Take: How a model describes humans affects how it cooperates with humans. This area is underrated in alignment work.
5. AI Software
1. OpenAI's new funding round targets a $1.25 trillion valuation
- Summary: OpenAI plans to raise about $30 billion in a new funding round, corresponding to a target valuation of about $1.25 trillion.
- Source: Bloomberg · 2026-10-23
- Editor's Take: A valuation running this far ahead of revenue relies on the self-reinforcing loop of the compute narrative. Until real money comes in, this number is more like a letter of intent.
2. Anthropic's IPO is imminent, early investors get a bump first
- Summary: Anthropic is expected to go public this year, with a valuation possibly reaching as high as $2 trillion. TechCrunch went through the pre-IPO paper value of some early investors' stakes.
- Source: TechCrunch · 2026-10-23
- Editor's Take: The people who make the most money are the ones who paid in first. This IPO will formally put AI's wealth effect onto the public market.
3. Agent cloud startup Daytona valued at $1 billion
- Summary: Daytona announced it closed a $60 million Series B led by FirstMark Growth, reaching a valuation of $1 billion. The company builds cloud environments for AI agents to run in.
- Source: TechCrunch · 2026-10-23
- Editor's Take: If agents are going to do work, they need somewhere to run, and it needs to be isolatable and rollback-able. Heat around this kind of infrastructure will rise along with agent deployment.
4. TechCrunch starts an AI graveyard list
- Summary: TechCrunch has begun maintaining an AI graveyard list, cataloging AI projects and startups that didn't make it. The first batch of entries includes Relay, a Zapier alternative that shut down completely in August.
- Source: TechCrunch · 2026-10-24
- Editor's Take: Nobody builds a cemetery in a bull market. Listing the projects that shut down is a mirror for those still raising.
5. Nvidia as "AI banker" faces its first stress test
- Summary: CNBC reports that as credit markets tighten around Nvidia's key partners, Nvidia's role as a financier of global AI buildout is facing its first major test, with neocloud stocks under pressure this week.
- Source: CNBC · 2026-10-24
- Editor's Take: Compute demand propped up by vendor credit doesn't show up on the income statement. Once credit tightens, demand shrinks with it — and this chain transmits faster than you'd think.
6. Don't let your AI bot log into your bank account
- Summary: CNET reminds users not to hand bank account information to agents. The piece names products like Meta Muse, OpenClaw, and OpenAI's dots, pointing out that once such authorization leaks, the chain of accountability is very long.
- Source: CNET · 2026-10-24
- Editor's Take: Letting an agent operate your account for you is like handing your password to a third party. Between authorization convenience and fund security, there's no mature middle layer yet.
Track Stats: Humanoid robots 3 · Autonomous driving 3 · World models and physical AI 2 · AI large models 6 · AI software 6, 20 total. The three physical-AI sections total 8 items, 40% of the total.
