WestTide · AI · 2026-11-16 · Issue 130
Today's Briefing: The aftershocks of Nvidia's leadership change are still spreading. CNBC calls Anthropic's IPO a make-or-break battle for the entire industry, and Sound Ventures' $7 billion Anthropic stake has become the fattest windfall of the year. The cybersecurity controversy around Chinese open-weight models has been laid out on the table by SCMP. The physical side is thin today — the runway-painting robots at Hong Kong Airport and the driverless trucks at a Tibetan mine are the rare "machines actually doing work" stories.
Editor's Note: News at the model layer increasingly reads like an earnings report; news at the physical layer increasingly reads like a construction site log. The latter isn't pretty, but it determines whether the former's valuations can actually land.
1. Humanoid Robots
1. Hong Kong Airport hands runway painting to robots
- Summary: Hong Kong International Airport has started using robots to repaint runway markings. This kind of work used to require overnight runway closures after flights stopped, with people standing on the runway pushing equipment along — safety and labor costs all crammed into those few nighttime hours. The airport's framing: "fresh coat, lighter load."
- Source: SCMP · 2026-11-15
- Editor's Take: Airport ground robots are an underrated business. Closed environments, fixed routes, regulators willing to approve — they'll turn a profit five years before humanoids do.
2. Sony's Aibo robot revolution looked like a dog
- Summary: This episode of The Verge's podcast looks back at Aibo. In the late '90s Sony was convinced the robot era was coming, and what it served up was a robot dog that wagged its tail and recognized people — selling companionship, not function. That generation didn't carry Sony's robot dreams, but it left behind a judgment: robots sell emotional value first, labor second.
- Source: The Verge · 2026-11-15
- Editor's Take: Twenty-odd years later, humanoid robot companies' launch events still use the same persuasion playbook — first make you think it's cute, then talk about what it can do.
3. Robot mowers rack up 255,000 miles
- Summary: Greenzie's autonomous mowers passed 255,000 miles cumulative last week. That number means nothing on a passenger car, but on a mower blade it's kind of interesting — they run parks, sports fields, and municipal green spaces. Slow, repetitive, day after day.
- Source: Electrek · 2026-11-15
- Editor's Take: Parks and lawns are where autonomy first made the business work. Slow speed, repetitive routes, affordable if something goes wrong — all three boxes checked.
2. Autonomous Driving
4. Chinese self-driving companies start looking for Nvidia backups
- Summary: This episode of TechCrunch Mobility is about the "Nvidia backup plan" in China's autonomous driving industry. After Jensen Huang stepped down on November 12, a batch of Chinese companies that had long pinned their compute on American chips have to redo the supply chain math.
- Source: TechCrunch · 2026-11-15
- Editor's Take: Swapping chips isn't swapping parts. Toolchains, quantization schemes, automotive certification all have to be rerun — that money usually takes two years to work through the financials.
5. Tesla sells Cybercab to individuals, $299/month subscription unavoidable
- Summary: Tesla opened invitation-only orders for the Cybercab on Saturday. The car is $29,990, and buyers must also subscribe to the $299/month Autonomy service. Fine print includes no tinting, no modifications, and the subscription price may change. The same day, Musk called YouTube creator Marques Brownlee a "chicken" on X — the two had a bet involving a head shave.
- Source: Electrek · 2026-11-15
- Editor's Take: They sell you the car and rent you the steering wheel by the month. That's software-company pricing, not automaker pricing.
6. High altitude, bitter cold, zero drivers: autonomous mining trucks take on a Tibetan mine
- Summary: Electrek reports autonomous mining trucks are now operating at a high-altitude mine in Tibet. Not being able to hire qualified equipment operators is a chronic problem for mines anyway, and with oxygen deprivation and extreme weather, fewer and fewer people are willing to go up the mountain. Fewer and fewer people willing to go up the mountain — so the zero-driver solution landed here.
- Source: Electrek · 2026-11-15
- Editor's Take: Mining trucks are the least sexy and most profitable scenario in autonomy. Twenty km/h, closed routes, trip after trip — no need to tell an urban-road story.
3. World Models · Physical AI
7. LeCun at CoRL: if you want human-level intelligence, stay away from LLMs
- Summary: Attendees relayed Yann LeCun's remarks at CoRL 2026. The core line: "If you're interested in human-level AI, don't work on LLMs." His alternative is the same old pitch — world models that can predict how the physical world evolves, rather than pushing next-token prediction to its limit.
- Source: Reddit · TheMachineLearning · 2026-11-15
- Editor's Take: He's been saying this since 2022. His position hasn't changed; what's changed is that fewer and fewer people are willing to listen. Lab money and attention are all bet on scaling parameters.
4. AI Large Models
8. EU's open-source model Europa hits 1 trillion parameters
- Summary: Italian company Domyn completed EUROPA's training trials on public supercomputers, with a configuration approaching 1 trillion parameters. EUROPA is the EU's own open-source model project, running on public compute rather than some company's data center.
- Source: maccus.fr · 2026-11-15
- Editor's Take: Europe doesn't lack compute for models; it lacks people willing to use them. Hitting 1T parameters won't win developers unless there are downstream vendors willing to rewrite their code for it.
9. JetBrains asked 12 models 1,200 times — they prefer PHP
- Summary: JetBrains ran 1,200 controlled tests across 12 models. The conclusion is a bit counterintuitive: models perform better on PHP than on Python. The tests covered the same set of programming tasks, only swapping languages.
- Source: JetBrains AI Blog · 2026-11-15
- Editor's Take: Models learn the internet's corpus distribution, not a language's merits. PHP has high Q&A density and uniform answer formats; Python's corpus is stuffed with half-finished notebooks.
10. Epoch AI updates that "most important number" again
- Summary: In its Gradient Updates newsletter, Epoch AI updated the long-tracked key AI metric. This time it's the first version on this curve with explained confidence intervals, and they also broke down a few commonly miscited statistical definitions.
- Source: Epoch AI · 2026-11-15
- Editor's Take: Every debate about whether AI has peaked ultimately comes back to this curve. Lots of people argue; few read the original methodology.
11. Chinese open-weight models put cybersecurity on the table
- Summary: SCMP's report asks a very practical question: can Chinese AI companies curb the cyberattack risks of open-weight models without curbing their own competitiveness along with it? Once weights are released they can't be recalled — the only constraints left are terms of service and downstream auditing.
- Source: SCMP · 2026-11-15
- Editor's Take: Safety clauses written into an agreement won't stop a weight file. This is an unavoidable cost of the open-source path — admit it early and save yourself trouble.
12. A lawyer upgrades an old 8×V100 server to run DeepSeek v4.1 Flash
- Summary: Tom's Hardware reports a lawyer upgraded his private 8-GPU V100 server to run about 80% of DeepSeek v4.1 Flash at 4-bit NVFP4 precision, adding 1TB of RAM and an RTX 3090, and fixed crashes caused by a bad PSU along the way.
- Source: Tom's Hardware · 2026-11-15
- Editor's Take: The value here isn't the model, it's the barrier to entry. Open-weight inference has dropped to the level where one person can get it running over a couple of weekends.
13. AI slop papers published under prominent academics' names
- Summary: Inside Higher Ed reports that a batch of low-profile academic journals published papers credited to prominent scholars, who deny involvement. One of them said the drafts were "extremely shallow, or simply generated."
- Source: Inside Higher Ed · 2026-11-13
- Editor's Take: Peer review at subscription journals was already loose; AI just tore the seam wider. The reputations of unwitting scholars get burned for it.
14. Most AI moats are just fuel
- Summary: adlrocha revisits his March judgment on Substack: intelligence itself is becoming a commodity, and the real moat is the context you feed the model. Half a year later his conclusion hasn't changed — just a few more pieces of evidence.
- Source: adlrocha · 2026-11-15
- Editor's Take: People who state "context is the moat" as a conclusion usually don't have much context themselves. The price of that statement being true is that you first need a batch of data others can't get.
5. AI Software
15. Anthropic's IPO treated as a physical for the whole industry
- Summary: CNBC says this IPO is a make-or-break battle for the tech industry. The backdrop is stacked: rising interest rates, a nationwide AI backlash, a political shift in Washington, plus Jensen Huang's sudden exit. The market needs a sample that proves the AI business adds up.
- Source: CNBC · 2026-11-15
- Editor's Take: One company's prospectus having to answer valuation questions for the entire industry itself shows the market has lost the ability to price independently.
16. Madonna's manager holds $7 billion via Anthropic stake
- Summary: Bloomberg reports that Sound Ventures, co-founded by Guy Oseary, holds an Anthropic stake worth about $7 billion — one of the biggest winners of this AI IPO wave. Oseary's day job is managing Madonna.
- Source: Bloomberg · 2026-11-15
- Editor's Take: The biggest returns in this AI wave didn't come from technical judgment — they came from whether you dared sign in those few months of 2022. No amount of detailed industry analysis beats one bet.
17. A $75,000-a-year AI school where half the students miss math targets
- Summary: Fortune reports on AI-driven private schools like Alpha School, with tuition up to $75,000 a year, backed by everyone from Silicon Valley to the Trump administration. Yet half the students at this school didn't meet the set math targets.
- Source: Fortune · 2026-11-14
- Editor's Take: AI teaching is best at looking "personalized" and worst at producing verifiable results. This business sells parents' anxiety, not test scores.
18. Apple bets smart glasses on Siri hooking into apps
- Summary: Bloomberg's Power On says Apple believes the killer feature of smart glasses is Siri being able to directly invoke third-party apps, not hardware specs. The same issue mentions progress on Face ID for the iPhone Duo.
- Source: Bloomberg · 2026-11-15
- Editor's Take: You can't cram compute into glasses temples, so Apple's plan is to offload the work to the phone and the cloud. That means gen one can only be an accessory, not a standalone device.
19. Hackers are already running attack chains with AI agents
- Summary: AWS's threat intelligence team observed an AI agent called ARTEX used in cybercrime operations, over an observation window from October 3 to November 5. The report focuses on something else: once humans are removed from the attack chain, the speed changes.
- Source: AWS Security Blog · 2026-11-15
- Editor's Take: Defenders are still talking about "AI-assisted analysts" while attackers have already taken the human out of the process. That generation gap will turn into concrete incidents within six months.
20. Top scorer on Korean tax exam admits to using AI to cheat
- Summary: Reuters reports the top scorer on South Korea's tax accountant exam admitted to using AI during the test. Exam authorities only started investigating after the score looked anomalous.
- Source: Reuters · 2026-11-15
- Editor's Take: Exams are the first institution AI breaks through, because they only verify results, not process.
21. Job seekers start pushing back on AI interviews
- Summary: TechRadar interviewed a batch of job seekers, many of whom view companies using AI for first-round interviews negatively. One person's exact words: "It would make me think twice about applying." Companies cite screening costs; candidates cite being treated as data points.
- Source: TechRadar · 2026-11-15
- Editor's Take: The interview costs companies save will eventually be deducted from the quality of their candidate pool. Especially noticeable for senior roles.
22. Torvalds wonders if Linux's AI patch flood has peaked
- Summary: The Register reports Linus Torvalds is considering whether AI-generated patches to the Linux kernel have passed their peak. The kernel community's biggest headache these past two years is review bandwidth — code was never the problem. Low-quality submissions keep piling in.
- Source: Slashdot / The Register · 2026-11-14
- Editor's Take: Maintainers' time is the kernel's scarcest resource. Patch volume going up tenfold doesn't make review capacity go up tenfold.
Track Statistics
| Track | Count |
|---|---|
| Humanoid Robots | 3 |
| Autonomous Driving | 3 |
| World Models · Physical AI | 1 |
| AI Large Models | 7 |
| AI Software | 8 |
| Total | 22 |
Physical Frontier | West Tide
Physix Frontier