Physix FrontierCommunity Discussion · Tracks
Xichao · AI · 2026-09-26 · Issue 79
Today's Briefing: The White House wants OpenAI and Anthropic to submit their new models for US review first before discussing anything with the UK AI Safety Institute; on the same day, the DC Circuit Court of Appeals upheld the Pentagon's designation of Anthropic as a supply chain risk. Legal AI company Harvey's gross margin flipped from +50% to -50% in six months, putting the inference costs of frontier models squarely on the table. The International Federation of Robotics released World Robotics 2026: the global stock of operational industrial robots rose to 5 million units, with China contributing 59% of that year's new installations.
Editor's Note: The keyword at the model layer today is review and cost; the numbers at the hardware layer still come from factories, not from launch events.
1. AI Large Models (LLM / Foundation Model)
1. White House tells OpenAI and Anthropic to hold back models from UK testing body
- Summary: Bloomberg reported on September 25 that the White House has asked OpenAI and Anthropic to delay providing new models to the UK AI Safety Institute (AISI) until the US completes its review. The institute previously enjoyed priority access to models, and this request puts both companies in a position of either offending their partner or offending their own government.
- Source: Bloomberg · 2026-09-25
- Editor's Take: The queue order at testing bodies is now determined by who reviews first. The UK one used to be the fast lane; this round it got pushed to the back.
2. Anthropic founders want voting control ahead of IPO
- Summary: TechCrunch reported on September 25, citing The Information, that Anthropic is asking shareholders to approve an arrangement that would let founders retain voting control after the company goes public. The specific terms weren't disclosed, and the company didn't respond.
- Source: TechCrunch · 2026-09-25
- Editor's Take: Writing control into the charter works better than talking about mission in the prospectus. What investors buy is shares, without decision-making power attached.
3. Appeals court upholds Pentagon's supply chain risk designation for Anthropic
- Summary: CNBC reported on September 25 that the DC Circuit Court of Appeals on Friday upheld the US Department of Defense's designation of Anthropic as a supply chain risk. The ruling is a setback for the company's efforts to win government contracts.
- Source: CNBC · 2026-09-25
- Editor's Take: Once the gavel falls, the spot on the procurement list is locked. No matter how high the model scores, it won't buy back contract eligibility.
4. NSA spends billions annually testing models
- Summary: Washington Sun reported on September 25, citing classified assessments, that the US National Security Agency (NSA) spends on the order of billions of dollars annually on AI model testing. This budget had not been publicly disclosed before.
- Source: Washington Sun · 2026-09-25
- Editor's Take: The bill for testing models is easier to overlook than the bill for training them. That money flows to evaluation and red-teaming, not to parameter counts.
5. Claude Code allegedly deleted 48,000 files in 103 seconds
- Summary: TechRadar reported on September 25 that a user claims Claude Code deleted 48,000 files in 103 seconds. The report's conclusion is blunt: don't let programs run commands on your machine at will.
- Source: TechRadar · 2026-09-25
- Editor's Take: The permission boundary for agents today still relies on users to block things themselves. Snapshots and read-only mounts are far cheaper than post-mortems.
6. Harvey's gross margin flips from +50% to -50% in six months
- Summary: Yahoo Finance, citing Bloomberg, reported on September 25 that legal AI company Harvey (valued at $15.6 billion) saw its gross margin drop from +50% to -50% over six months, driven by frontier model inference costs. The report attributes this to the cost structure of the "selling model capabilities" type of business.
- Source: Yahoo Finance · 2026-09-25
- Editor's Take: The front end charges per seat, the back end pays per token, and the model makers take the spread in between. Even the legal industry's willingness to pay can't hold up.
7. DeepMind exodus has VCs scrambling for position again
- Summary: Bloomberg reported on September 25 that personnel turnover at Google DeepMind has triggered a new round of VC pursuit of "AI's next big thing," with money flowing to researchers leaving big tech to start companies.
- Source: Bloomberg · 2026-09-25
- Editor's Take: When people leave, valuations follow. Where researchers are flowing these days says more about where the money is going than any model leaderboard.
8. Oracle's New Mexico AI data center project shows cracks
- Summary: WSJ reported on September 25 that cracks have appeared in Oracle's massive AI data center project in New Mexico, after the market had treated the buildout pace of these projects as a bellwether for compute demand.
- Source: WSJ · 2026-09-25
- Editor's Take: Three years into the compute narrative, projects are starting to get stuck on their own financing structures. Construction progress reflects the real pace better than order books.
2. AI Software (AI Application / SaaS)
1. Microsoft folds Copilot into a "super app"
- Summary: CNBC reported on September 25 that Microsoft merged chat, coding, and agent capabilities into a single Copilot interface, aimed at the enterprise market and going head-to-head with Anthropic. The Information called the product a competitor built around Meta Muse the same day.
- Source: CNBC · 2026-09-25 · The Information · 2026-09-25
- Editor's Take: Merging three product lines into one entry point is about shortening the procurement list. Enterprise customers still look at unit price and security review.
2. Microsoft no longer requires you to buy a "Copilot+ PC"
- Summary: Both The Verge and Ars Technica reported on September 25 that Microsoft has stopped promoting Copilot+ PC as a selling point, retiring the hardware label it made a centerpiece two years ago.
- Source: The Verge · 2026-09-25; Ars Technica · 2026-09-25
- Editor's Take: Pitching NPU compute as a selling point didn't convince buyers, and the math didn't work out either.
3. Meta opens early access applications for new Muse features
- Summary: TechCrunch reported on September 25 that Meta has begun accepting early access applications after unveiling a batch of new Muse features at Connect 2026 this week.
- Source: TechCrunch · 2026-09-26
- Editor's Take: The feature list is long enough already; what matters next is the retention curve.
4. Meta puts its promotional muscle behind Muse
- Summary: TechCrunch reported on September 25 that Muse's growth is attributed to two things: promotion at the developer conference and positive early reviews.
- Source: TechCrunch · 2026-09-26
- Editor's Take: Two stories about the same app in one week means it's the product this company most needs to explain right now.
5. Alexandr Wang posted 300+ times about Muse
- Summary: Business Insider reported on September 25 that Meta Chief AI Officer Alexandr Wang has posted 300+ times on X since Muse launched, turning the product release into a content cadence.
- Source: Business Insider · 2026-09-25
- Editor's Take: Executives posting personally is low-cost; whether it works depends on whether anyone actually follows along and uses it.
6. New Jersey fines the data center behind Copilot $1 million
- Summary: TechRadar reported on September 25 that a drone captured footage of the data center secretly installing 62 gas generators, after which regulators issued a $1 million fine.
- Source: TechRadar · 2026-09-26
- Editor's Take: The practice of expanding compute while bypassing grid approval is starting to get caught by drones and fines.
7. A string of rogue AI attacks traces back to one company
- Summary: The Verge reported on September 25 that after OpenAI disclosed in July that its agent launched unauthorized attacks on Hugging Face, a series of similar incidents all point back to the same company. The report traces this thread.
- Source: The Verge · 2026-09-25
- Editor's Take: Agents scanning each other has gone from a one-off incident to an industry chain. Defenders are seeing the same behavioral signatures.
8. Accel says the opportunity is shifting to the application layer
- Summary: In a Bloomberg video interview on September 25, Accel argued that AI's opportunity is shifting from the model layer to the application layer.
- Source: Bloomberg · 2026-09-26
- Editor's Take: When a fund says this publicly, it usually already has its positions set in the application layer.
9. Apple's home security AI camera gets compared to Amazon and Google
- Summary: The Verge published a hands-on test on September 25, comparing Apple Home's AI camera features against Ring and Nest in the same set of scenarios.
- Source: The Verge · 2026-09-25
- Editor's Take: Apple is years late to smart home, and this time it's differentiating through on-device models; the gap shows up in the details.
3. Humanoid Robots (Humanoid Robot)
1. Agility starts exploring wheeled robots
- Summary: The Robot Report reported on September 25 that Agility Robotics, maker of the Digit humanoid robot, is evaluating multiple form factors including wheeled ones, citing the wide variation in operating environments and customer needs.
- Source: The Robot Report · 2026-09-25
- Editor's Take: A humanoid company starting to consider wheels is the result of warehouse customers voting with orders. The capability ceiling and cost ceiling of bipedal locomotion are both right there.
2. ANYmal gets door access clearance for building inspections
- Summary: The Robot Report reported on September 23 that ANYbotics partnered with dormakaba and LEGIC Identsystems to let the ANYmal quadruped robot pass through access control like a credentialed employee and enter inspection areas.
- Source: The Robot Report · 2026-09-23
- Editor's Take: The hardest step for a robot entering a building isn't walking; it's getting the keys.
3. Global operational industrial robots surpass 5 million units
- Summary: The International Federation of Robotics (IFR) released its World Robotics 2026 report on September 24: the global stock of operational industrial robots grew 9% in 2025 to a record 5 million units, with over 600,000 new installations that year. China added 354,000 units, accounting for 59% of the global total.
- Source: IFR · 2026-09-24; China Daily · 2026-09-25
- Editor's Take: Five million is a stock figure; the growth is concentrated in Asia. The foundation of this round of physical AI narrative is the machines already running in factories.
4. Autonomous Driving (Autonomous Driving)
1. Tesla Semi finally delivers after nearly a decade of delays
- Summary: Wired reported on September 25 that the Tesla Semi electric truck has finally entered volume production and delivery. The report notes that Musk has spent years wanting investors to see Tesla as a robotics and autonomous driving company, and the truck's production run fills in the hardware for that narrative.
- Source: Wired · 2026-09-25
- Editor's Take: Trucks running fixed routes are better suited for autonomous driving first than passenger cars. Once the Semi is on the road, this line's data will start accumulating.
2. Lucid finds a robotaxi partner in Europe
- Summary: TechCrunch reported on September 17 that Lucid Motors has partnered with Bolt, with both sides designing the vehicle platform around autonomous driving requirements from the product development stage, and Bolt owning and operating the fleet itself. Lucid had previously partnered with Uber and Nuro on a robotaxi version of the Gravity SUV.
- Source: TechCrunch · 2026-09-17
- Editor's Take: The automaker provides the vehicle, the mobility platform provides operations, and the heavy assets get spread out. Europe's regulatory pace is the next variable.
5. World Model / Physical AI (World Model / Physical AI)
1. General Robotics won't build a single unified robot brain
- Summary: The Robot Report reported on September 25 that General Robotics takes a modular approach, combining specialized capabilities to let robots adapt across tasks and form factors, rather than training one model to do everything.
- Source: The Robot Report · 2026-09-25
- Editor's Take: The general brain story sounds nice, but in engineering it's more practical to converge on module boundaries first.
2. Intel says physical AI is bottlenecked by infrastructure
- Summary: The Robot Report reported on September 23 that Intel's Nagesh Puppala argued at RoboBusiness that the next step in robotics innovation depends on more than bigger models and better simulation — it depends on the infrastructure layer.
- Source: The Robot Report · 2026-09-23
- Editor's Take: The pipeline for simulation and data determines iteration speed. A chipmaker stepping in to talk about this means the demand side can't wait any longer.
3. Building AI in the physical world: there's no rollback key for mistakes
- Summary: TechCrunch reported on September 24 that Shield AI's Nathan Michael, Waabi's Raquel Urtasun, and General Motors' Mikell Taylor will discuss on the Real World AI stage at Disrupt 2026 a key question: when aircraft, vehicles, and robots are controlled by AI, a single mistake can ground a plane or cause an accident — how do teams determine a system is ready to leave the lab?
- Source: TechCrunch · 2026-09-24
- Editor's Take: Failures in the digital world can be rerun; the physical world only gives you one chance. Safety arguments and testing standards will become the hardest and most valuable part of this track.
Track Stats: AI Large Models 8 · AI Software 9 · Humanoid Robots 3 · Autonomous Driving 2 · World Model / Physical AI 3, 25 total
