Physix FrontierWestTide · AI · 2026-10-31 · Issue 114
Today's Briefing: Nvidia is in talks to provide credit backing for a cloud provider's $45 billion Anthropic order — the circular financing chain has stretched from Silicon Valley to Asia. DoorDash admits external AI agents have only placed a few hundred orders to date, and Amazon's CEO says shopping agents are "still not great." On the humanoid robot front, Dreame's home robot has started sorting laundry by fabric, while academia is laying the groundwork for dexterous manipulation and locomotion control.
Editor's Note: The numbers worth looking at today are all small — a few hundred orders, a few thousand pages, a few dozen seconds of queuing. The volume of the industry narrative and these numbers are often not on the same scale.
1. Humanoid Robots
1. Dreame's home robot starts sorting laundry by fabric, founder is 29
- Summary: Bloomberg reports that Dreame's home robot Echo S1 debuted at a Berlin trade show in September and can separate clothes by fabric and color. The company is led by a 29-year-old founder, with a product line stretching from robot vacuums all the way to household robots.
- Source: Bloomberg · 2026-10-29
- Editor's Take: Sorting is the easiest part of housework to demo. Folding, putting away, and returning things to the cabinet is the part you have to repeat twenty times a day.
2. DexSuite: one framework connecting robot arms, dexterous hands, and tasks
- Summary: DexSuite, submitted to arXiv on October 27, brings different robot arms, different dexterous hands, and different manipulation tasks into a single framework, aiming to spare dexterous manipulation experiments from rebuilding the environment every time.
- Source: arXiv · 2026-10-30
- Editor's Take: The bottleneck in dexterous manipulation is whether testbeds can be compared across the board — the algorithms have long been good enough.
3. FORGE: the online reference generation that gets humanoid robots running
- Summary: FORGE, from the same batch of papers, does two things: generate walking reference trajectories online, then have a whole-body controller follow them. The goal is to make "motions you can imagine" line up with "motions the joints can actually follow."
- Source: arXiv · 2026-10-30
- Editor's Take: Watch enough running and jumping videos and it's easy to forget — the torque at landing is what shows up on the hardware bill.
4. Retargeting human motion onto humanoid arms, the error shows up at the wrist
- Summary: An empirical study dissects motion retargeting for humanoid arms and finds that how wrist targets are constructed and how joint angle limits are set directly determines whether the retargeted motion is still usable.
- Source: arXiv · 2026-10-30
- Editor's Take: Half the price of teleoperation data goes to these angle limits that nobody wants to write into the abstract.
2. Intelligent Driving
1. Rivian's autonomous driving bet faces its first real test
- Summary: The Verge reports that in Rivian's one-hour Q3 earnings call, CEO RJ Scaringe spent less time on R2 production ramp than on the autonomous driving roadmap. The company's point-to-point robotaxi plan is starting to meet reality.
- Source: The Verge · 2026-10-31
- Editor's Take: When production capacity and autonomous driving are competing for the same money, how a call allocates its minutes is more honest than any slide deck.
2. Tesla swaps referral rewards for FSD
- Summary: Tesla launched a new referral program in the US: the referred party gets one month of free FSD, the referrer gets FSD usage credits, and the old car-purchase rewards are gone.
- Source: Electrek · 2026-10-31
- Editor's Take: Trading software credits for sales is essentially booking customer acquisition cost against future gross margin.
3. Waymo's fleet is hogging public fast chargers
- Summary: InsideEVs reports that many public fast-charging stations in California are being occupied for long stretches by Waymo robotaxis, lengthening queue times for ordinary EV owners, with complaints concentrated in recent months.
- Source: InsideEVs · 2026-10-31
- Editor's Take: In the robotaxi cost sheet, charging queue time has never been counted into passengers' waiting time.
4. GM pays $220 million to settle driving data lawsuit
- Summary: Oregon Attorney General Dan Rayfield announced that GM will pay $220 million (about 1.47 billion yuan) to settle a joint lawsuit brought by 36 states over the sale of Smart Driver driving data.
- Source: Ars Technica · 2026-10-30
- Editor's Take: When the data was sold, it was priced per record; when it was bought back, it was priced per state.
5. Hong Kong ride-hailing license deadline closes; Uber, Tada, Amap, and DiDi all applied
- Summary: SCMP reports that before the Hong Kong ride-hailing license application deadline, at least 4 major operators confirmed they had submitted, including Uber, Tada, Amap, and DiDi.
- Source: SCMP · 2026-10-30
- Editor's Take: Licenses pull the gray zone into regulation; next comes the fight over fleet compliance costs.
3. World Models · Physical AI
1. The more planning steps you add, the worse agents in world models get
- Summary: The October 27 paper "When More Planning Hurts" evaluates learned planners for world models and finds that deeper planning does not reliably bring gains — on some tasks performance actually reverses.
- Source: arXiv · 2026-10-30
- Editor's Take: When the world model's resolution isn't high enough, computing a few more steps just amplifies the error a few more times.
2. Millimeter-wave radar can grow a 3D occupancy-language field too
- Summary: The same batch of papers uses mmWave radar to build a 3D occupancy-language field, giving radar point clouds semantics, aiming to fill in perception in rain, fog, and at night where cameras struggle.
- Source: arXiv · 2026-10-30
- Editor's Take: Now that lidar has gotten cheaper, mmWave's value returns to the weather it's best at.
3. π0.5 on BEHAVIOR-1K: which inference settings actually help
- Summary: A paired evaluation compares various inference settings of π0.5 on BEHAVIOR-1K item by item, trying to sort out which parameters really affect success rates on long-horizon household tasks.
- Source: arXiv · 2026-10-30
- Editor's Take: Half the reproducibility problem for embodied models comes down to nobody writing down what the inference settings actually were.
4. AI Large Models
1. Nvidia to provide credit backing for a $45 billion Anthropic order
- Summary: The Information reports that Nvidia is in talks to provide billions of dollars in credit support for a cloud provider's $45 billion contract with Anthropic — the money to buy compute is backstopped by the seller's ecosystem.
- Source: The Information · 2026-10-30
- Editor's Take: Chipmakers are now guaranteeing their customers' customers — revenue recognition and risk recognition get split across two companies' books.
2. After reading 300-plus pages of Anthropic's prospectus, two questions remain
- Summary: CNBC's breakdown says Anthropic's S-1 runs over 300 pages, and after disclosure the market still cares most about revenue structure and compute commitments — these two determine the post-IPO valuation anchor.
- Source: CNBC · 2026-10-30
- Editor's Take: A prospectus can spell out how the losses are composed; it can't spell out whether the compute contracts three years out will still be worth this price.
3. Circular AI deals spread to Asia
- Summary: Bloomberg reports that circular financing structures in AI are starting to appear in Asian deals, with investors and investees as each other's customers, and investors worry risk transmits along this chain.
- Source: Bloomberg · 2026-10-30
- Editor's Take: The same money gets booked as revenue once on each of two companies' statements — auditors can see it, cash flow can't.
4. The patch gap for open-weight models
- Summary: Epoch AI published research: Z.ai's open-weight model GLM 5.3, released in August, can identify and exploit vulnerabilities in recent software versions. Between open models and the patch cadence, a time gap has opened up.
- Source: Epoch AI · 2026-10-30
- Editor's Take: What open-sourcing releases is capability — and also an instruction manual for exploitation.
5. AI-written code still frequently fails security scans
- Summary: An evaluation cited by TechRadar shows AI-generated code still has an unsatisfactory pass rate on security checks, with Claude and GPT-6 topping the problem leaderboard; the report calls progress "slow."
- Source: TechRadar · 2026-10-31
- Editor's Take: Code generation long ago outran code review — and the security team ends up paying the bill.
6. Altman sympathizes with mathematicians, but the math offensive continues
- Summary: In a Cheeky Pint podcast interview, OpenAI CEO Sam Altman said he understands mathematicians' frustration, while making clear OpenAI will keep pushing results in math and physics.
- Source: The Verge · 2026-10-30
- Editor's Take: Publish results first, then ask people to verify — the academic community's review cadence has been treated as an appendix to the release process.
5. AI Software
1. DoorDash says external AI agents have only placed "a few hundred orders"
- Summary: DoorDash disclosed on its earnings call that orders from external AI agents total only a few hundred so far — a stark gap from the buzz about "agents taking over e-commerce."
- Source: TechCrunch · 2026-10-30
- Editor's Take: The barrier to agents placing orders isn't the model — it's payment authorization and who eats the refunds.
2. Starbucks wants to show you a personalized menu at the drive-thru
- Summary: Starbucks' CEO mentioned on the earnings call that drive-thrus may in the future display personalized digital menus based on user habits; launch timing and store scope have not been announced.
- Source: The Verge · 2026-10-31
- Editor's Take: Recommender systems have entered the drive-thru lane — in those thirty seconds of queuing, the only thing that can really change is the add-ons.
3. AI law firm Crosby bids for Weil Gotshal's assets
- Summary: TechCrunch reports that AI law firm Crosby has bid for the assets of the disintegrating Wall Street stalwart Weil, Gotshal & Manges; the legal industry sees the deal as a case study of AI legal services replacing the traditional billing model.
- Source: TechCrunch · 2026-10-31
- Editor's Take: Buying a law firm's assets is easy; buying its clients' trust is far harder.
4. Amazon CEO: shopping agents are "still not great"
- Summary: Andy Jassy said in an interview he isn't worried about shopping agents and crawlers bringing down Amazon's site, on the grounds that the experience of using such agents is currently "still not great."
- Source: TechRadar · 2026-10-30
- Editor's Take: What the platform worries about is the day agents suddenly become good enough.
5. Granola adds dictation, going straight at Wispr
- Summary: Granola, which started with meeting notes, has launched a dictation feature, entering the voice-input arena and competing head-on with Wispr.
- Source: TechCrunch · 2026-10-30
- Editor's Take: Meeting notes and dictation share one audio pipeline — the product boundary was never solid to begin with.
6. Zest's agent books you a table over text
- Summary: Restaurant discovery app Zest launched an agent that helps users find restaurants and complete reservations via text message; the company sees it as the first version of a world where "the app is no longer the main entry point."
- Source: TechCrunch · 2026-10-31
- Editor's Take: Tasks with a clear success criterion, like booking a table, are where agents can most easily prove themselves.
7. Cisco has a 27-year-old engineer rewrite 98% of legacy code with unlimited tokens
- Summary: Tom's Hardware reports that Cisco handed the task of rewriting 98% of its legacy code to a 27-year-old engineer, equipped with unlimited AI tokens, over a 24-month period.
- Source: Tom's Hardware · 2026-10-30
- Editor's Take: Betting the rewrite on one person and a pile of tokens is a bet that the model can read the entire codebase.
Track tally: Humanoid robots 4 · Intelligent driving 5 · World models 3 · Large models 6 · Software 7 — 25 items total.
