WestTide · AI · 2026-11-21 · Issue 135
Today's Briefing: On its second day of trading, Anthropic's stock reversed course, and Reuters framed this cooldown as a "reality check" for the AI trade. In the same week, the Riemann Hypothesis—open for 167 years—was declared proven twice: once with Claude's involvement, and once by a human working alone. On the physical side, 21 humanoid robots finished a half marathon, while the Waymo–Uber partnership is winding down. The software layer is starting to set rules: Signal blocks screenshots by default, and several major US banks are putting the agent's bill back on users.
Editor's Note: Valuations can be repriced in a day; accountability gets settled over years. Money moves faster than the rules—by November, that's no longer a new topic.
1. Humanoid Robots
1. 21 humanoid robots finished a half marathon—a year ago they were still falling over
- Summary: A TechRadar reporter ran a half marathon alongside 21 humanoid robots. One line in the report is more worth remembering than the results: within a year, this batch of machines went from stumbling walks to breaking records.
- Source: TechRadar · 2026-11-20
- Editor's Take: A half marathon tests joint heat dissipation and gait robustness, not intelligence. Finishing means the motors and reducers have passed their test; being able to tighten a screw afterward is the next hurdle.
2. Letting legged robots see the path themselves: vision-language reasoning hooked up to a learned controller
- Summary: A new paper connects the reasoning output of a vision-language model to a learned low-level controller. In the field, legged robots first judge the nature of the terrain before deciding where to step, without needing to build a precise map first.
- Source: arXiv 2611.25215 · 2026-11-18
- Editor's Take: The value of this kind of work isn't in the demo video. Putting "understanding what it sees" and "walking steadily" into the same loop—that interface layer used to be hand-written by engineers.
3. Old-school planners are sometimes more reliable than learned policies
- Summary: The paper pits classical planners against learned policies on the same batch of tasks. The conclusion: the former is more stable in some scenarios, and the advantages of learned approaches aren't uniform.
- Source: arXiv 2611.25138 · 2026-11-18
- Editor's Take: This conclusion will make end-to-end people uncomfortable. But engineering choices are about failure modes, not average scores.
4. Robots need to learn to spend a whole day in a crowd
- Summary: Inhabitant breaks long-horizon human-scene interaction into subtasks with spatial constraints. The agent works continuously in a room where people are moving around, without resetting the environment at every step.
- Source: arXiv 2611.25005 · 2026-11-18
- Editor's Take: The hard part of home scenarios is disturbing others and being disturbed by them. Being willing to tackle long horizons amounts to admitting that single-step success rate is no longer the bottleneck.
2. Intelligent Driving
1. The Waymo–Uber partnership is winding down
- Summary: The Verge walked through the ins and outs of this breakup. On the August earnings call, Uber's Dara Khosrowshahi still described the partnership with Waymo as "very strong." A few months later, the fleets started building their own platforms.
- Source: The Verge · 2026-11-20
- Editor's Take: Platforms want to hold the traffic gateway; fleets want to hold the passengers. When they can't agree on who funnels users to whom, they just do it themselves.
2. Tesla's assisted driving gets approval in Italy, the EU's 12th country
- Summary: Italy's Ministry of Transport issued temporary type approval for Tesla's assisted driving. In Europe, the name has already changed from "Full Self-Driving" to "assisted driving." Italy is the 12th European country to give the green light.
- Source: Electrek · 2026-11-20
- Editor's Take: Southern Europe's narrow roads, roundabouts, and unmarked old towns are the real test. The approvals from the Nordic highway countries have limited reference value.
3. BYD makes a steering wheel that can be folded away
- Summary: InsideEVs got a look at the cabin of the Fang Cheng Bao Tai 9: the steering wheel folds away, and once retracted, the front row becomes one flat space. The design is meant for the time when "the car drives itself."
- Source: InsideEVs · 2026-11-21
- Editor's Take: The wheel can fold, but regulations may not allow it. Building it physically first is both a stance and a question posed to regulators.
3. World Models · Physical AI
1. MAWAM lets multiple robotic arms share one world-action model
- Summary: The paper proposes MAWAM, which builds collaborative manipulation into a shared world-action model—predicting the future while outputting actions. Tasks like joint carrying no longer rely on a central scheduler to assign order.
- Source: arXiv 2611.25105 · 2026-11-18
- Editor's Take: The bottleneck in multi-robot coordination has always been "who moves first." Stuffing prediction and decision-making into the same model amounts to letting the machines anticipate each other.
2. GEMS adds a layer of experiential memory to world models
- Summary: GEMS stores subgoals it has tried. When a long-horizon task reaches a fork, it checks memory first before letting the model generate the next stretch of imagination, so planning success no longer depends entirely on piling up samples.
- Source: arXiv 2611.25134 · 2026-11-18
- Editor's Take: The most expensive thing in a world model is imagination. Adding memory saves compute and keeps it from falling into the same pit twice.
3. Verify the plan before acting: Verify2Act stuffs a critic into latent space
- Summary: After a language instruction is turned into a manipulation plan, Verify2Act uses a latent-space world model and a critic to validate it before actual execution, intercepting plans that would obviously fail.
- Source: arXiv 2611.25243 · 2026-11-18
- Editor's Take: The cost of a robotic arm crashing once is far higher than a few extra seconds of compute. Verify-then-act should have been the default all along.
4. Identity recognition moves into an imagined latent space
- Summary: Identity World Model uses a world model to predict what the same person looks like under different viewpoints, occlusions, and lighting, then compares identities in latent space. Changes like switching jackets or looking down are no longer fatal.
- Source: arXiv 2611.25204 · 2026-11-18
- Editor's Take: The security market never lacked recognition rates; what it lacks is "still recognizes you after you change your jacket." The world model's job here is filling in viewpoints.
5. Weather models forecast better and better; researchers start auditing what they actually learned
- Summary: The paper audits a learned atmospheric dynamics system, pinpointing which biases come from the model's own approximations and which from the data, adding a layer of interpretable checks to black-box forecasting.
- Source: arXiv 2611.25119 · 2026-11-18
- Editor's Take: Accurate forecasts and understanding physics are two different things. When extreme weather really hits, the former's good luck won't hold up for long.
4. AI Large Models
1. The Riemann Hypothesis was declared proven twice this week—once with AI, once without
- Summary: Anthropic mathematician Levent Alpöge announced that, working with Claude, he had completed a proof of the Riemann Hypothesis, open for 167 years. His company had just become the first frontier lab to go public this week. Almost simultaneously, another mathematician posted his own proof. Neither draft has passed peer review.
- Source: TechCrunch · 2026-11-20
- Editor's Take: The math community's problem now is a shortage of reviewers. To really determine who's right, someone has to read through the mountain of formalized code the machine produced—and that job may itself be harder than the proof.
2. On day two of trading, the stock price flipped over the AI trade's hole card
- Summary: Reuters puts this pullback into a bigger picture. After a record-breaking IPO, the market didn't keep buying; the AI sector pulled back along with it, and investment banks and retail investors alike found themselves in the same boat.
- Source: Reuters · 2026-11-20
- Editor's Take: A valuation can ride a story for a day, not a quarter. The real pricing power sits on the compute-bill side.
3. Cursor tells lawmakers it's considering dropping Chinese models
- Summary: CNBC reports that Cursor, the coding platform under SpaceX, told Congress it is evaluating phasing out Chinese-made models, citing "geopolitical risk."
- Source: CNBC · 2026-11-20
- Editor's Take: Swapping models isn't hard; what's hard is how many benchmark points you lose afterward. The biggest buyers of open weights have always been these overseas application companies.
4. OpenAI writes a dedicated post on adversarial distillation
- Summary: OpenAI published a dedicated piece on how to deal with outsiders repeatedly training small models on its model outputs to siphon off capabilities, laying out detection and countermeasures.
- Source: OpenAI · 2026-11-20
- Editor's Take: The stronger the model, the more its outputs look like an asset. This line of defense will probably end up in watermarks and API terms—it can't be plugged technically.
5. Nvidia to buy a diffusion model team for over $500 million
- Summary: The Information reports that Nvidia is close to acquiring an AI company doing diffusion models for over $500 million. Chinese media say the target is Inception, which focuses on inference speed.
- Source: The Information · 2026-11-20
- Editor's Take: Buying a team is cheaper than buying cards. Nvidia is filling in the inference-speed link—once the cards are sold, it hopes the software won't drag either.
6. A US AI company openly discusses how it uses Chinese open weights
- Summary: Bloomberg brought in the head of a US AI company to explain how they plugged Chinese open-weight models into their products, laying out both the cost structure and the usage boundaries.
- Source: Bloomberg · 2026-11-20
- Editor's Take: Talk decoupling, but the books talk cost-effectiveness. The biggest benefit of open weights is that they can't be cut off—that matters more than benchmark scores.
7. METR updates the most misread curve
- Summary: METR released its latest report. The conclusion: the length of tasks models can complete is still rising fast—it's just that this curve has always been cited the wrong way.
- Source: MIT Technology Review · 2026-11-20
- Editor's Take: This chart is the most cited and the most easily mistaken for a conclusion. Before looking at the slope, look clearly at what the vertical axis actually measures.
8. Novartis CEO puts AI's payoff in terms of time
- Summary: The head of Novartis says AI can compress R&D cycles by years in target screening and clinical design—the value isn't in saving a few points of cost.
- Source: Bloomberg · 2026-11-20
- Editor's Take: Pharma has more patience with AI than Silicon Valley because it already has to wait ten years. Slow industries are actually more willing to bet.
5. AI Software
1. Signal blocks screenshots by default, citing AI agents peeking at screens
- Summary: Signal turns off screenshots by default and calls for a simple opt-out switch for developers at the system level, so screen-reading agents don't casually make off with private conversations.
- Source: Engadget · 2026-11-21
- Editor's Take: Agents on phones do their work by reading the screen—this cuts at the interface. More apps will follow suit.
2. Several major US banks say everything an agent does counts as authorized by you
- Summary: Several US banks have written into their terms of service that operations a user has an agent perform are treated as authorized by the user themselves; if phished or if the agent makes a mistake, the user bears the loss.
- Source: TechRadar · 2026-11-20
- Editor's Take: This disclaimer will go into every bank's service agreement. Users think they hired an assistant; legally, they hired themselves.
3. Cognition in talks for a new round at a $75 billion valuation
- Summary: Bloomberg reports that Cognition, which makes coding agents, is in talks for a new funding round at a $75 billion valuation.
- Source: Bloomberg · 2026-11-20
- Editor's Take: Coding is currently the only agent scenario proven to bring in money. However absurd the valuation, at least there's cash flow holding it up.
4. Google to shut down Assistant on smart speakers next year
- Summary: Google announced that starting next year it will discontinue Google Assistant on smart speakers, take the Nest app offline too, and fold the voice entry point entirely into Gemini.
- Source: The Verge · 2026-11-21
- Editor's Take: Changing a wake word is easy; changing a house full of already-installed devices is hard. The migration cost ultimately lands on old users.
5. Salesforce and Microsoft cook up a Copilotforce
- Summary: The two announced they're connecting Copilot with Salesforce's customer data, branding the partnership Copilotforce, focused on calling agents directly on your own business data.
- Source: TechRadar · 2026-11-20
- Editor's Take: Two companies that both make money selling seats teaming up shows that the marginal value of selling agents alone is declining.
6. Chinese attackers let agents find the vulnerability and break in themselves
- Summary: The Register reports that a group of China-linked attackers put agents into a target network, letting them autonomously discover a zero-day and complete the exploit, ultimately taking over a university server.
- Source: The Register · 2026-11-20
- Editor's Take: Vulnerability discovery has gone from person-months to machine-hours. Defenders are still patching quarterly—that pace gap is the entire source of the attack surface.
7. Google employees sign a petition to keep an AI-free search for schools
- Summary: Hundreds of Google employees signed a petition asking the company not to stuff AI into every search entry point, and to keep at least one clean version for schools.
- Source: Bloomberg · 2026-11-20
- Editor's Take: Insiders opposing their own product usually means they've seen a few metrics outsiders can't.
Track Stats: Humanoid Robots 4 · Intelligent Driving 3 · World Models & Physical AI 5 · AI Large Models 8 · AI Software 7, 27 items total, with the three physical sections making up 44%.
Physical Frontier | WestTide
Physix Frontier