Community Discussion · General

WestTide · AI · 2026-11-15 · Issue 129

West TideWest Tide4d ago2026/11/14 270 views

Today's Briefing: Anthropic's roadshow has become Wall Street's pricing litmus test this week, with seven co-founders, Jamie Dimon, and the compute bill all laid out on the table. Tesla is selling the Cybercab to individuals for $29,990, but skip the $299/month subscription and the car won't move. On the overseas hardware side, only two hard news items today: a military robotics company can't make payroll, and DeepMind open-sourced the datasets behind its weather models.

Editor's Note: Today's pool is 90% software and large models—the robotics section is eerily empty. The real signal isn't at any launch event; it's on Anthropic's roadshow deck: a model company is starting to let the public market price compute, and from now on it has to report quarterly.


1. Humanoid Robots

1. The military robotics company backed by Eric Trump didn't make payroll last month

  • Summary: Foundation Future Industries failed to pay employees on time last month as its cash ran dry. Records reviewed by CNBC show the founder had borrowed $4.6 million from the company to buy a house. The humanoid military robotics company counts Eric Trump as an advisor and investor; co-founder Mike LeBlanc said on Fox News just days ago that most combat would be handed to machines within 5 to 10 years.
  • Source: CNBC · 2026-11-14
  • Editor's Take: Military robotics never lacks order-book imagination—what it lacks is cash flow. Borrowing $4.6M to buy a house says more about this company's real scale than any battlefield demo.

2. Smart Driving

1. Tesla sells the Cybercab to individuals: $29,990, plus $299/month

  • Summary: Tesla opened Cybercab pre-orders Friday night to "select invited customers": the car is $29,990, the self-driving subscription is $299/month, and use is limited to Robotaxi operating zones. First deliveries are expected in December.
  • Source: TechCrunch · 2026-11-14
  • Editor's Take: Sell you the car first, then charge monthly for "it can drive itself." The car price is the entry ticket; $299 is the actual business.

2. No subscription, no movement: the Cybercab contract terms get dug up

  • Summary: Electrek obtained Tesla's self-driving subscription agreement, whose terms state that vehicles without an Autonomy subscription cannot enable self-driving features. In other words, if the owner stops paying, the car stays put.
  • Source: Electrek · 2026-11-14
  • Editor's Take: Ownership and usage rights are hard-split into two separate bills. Whether "you bought the car but can't drive it" holds up in contract law will fuel a pile of lawsuits over the next two years.

3. World Models · Physical AI

1. DeepMind open-sourced the datasets behind its weather models

  • Summary: DeepMind released two datasets, PARDIG and SatWeave, covering meteorology, hydrology and related scenarios for researchers to train and validate AI weather models. Such data used to be locked inside weather agencies and labs.
  • Source: Google DeepMind · 2026-11-14
  • Editor's Take: Open-sourcing models is a gesture; open-sourcing data is actually ceding ground. Where the reanalysis data and raw satellite observations are is where this niche's barrier to entry is.

2. 3D-printed figure blind-eval leaderboard: humans vote on models' spatial consistency

  • Summary: The community project Wojak-Bench 3D has different models generate 3D-printable outputs from the same sketch, and users blind-vote for the one that looks more alike. Model names are hidden before voting, and each browser gets one vote per pair.
  • Source: Wojak-Bench · 2026-11-15
  • Editor's Take: Human preference voting can't fix geometry errors, but it can fix "describe-the-picture" evaluation inflation. This crude approach is often closer to real perception than the FID table in a paper.

4. AI Large Models

1. Anthropic's IPO becomes a public pricing of AI assets

  • Summary: Bloomberg's tech show treated Anthropic's upcoming IPO as the week's key observation point, discussing how much premium investors are willing to pay for a model company still burning huge sums on compute.
  • Source: Bloomberg · 2026-11-14
  • Editor's Take: The AI bubble debate has raged for three years, and it all comes down to the subscription multiple on a prospectus. That's harder than any benchmark.

2. Who are Anthropic's seven co-founders?

  • Summary: CNBC profiled the seven people who founded Anthropic in 2021: four researchers, a former journalist, a public policy maker, and an engineer. Six years later, the company stands at the IPO door.
  • Source: CNBC · 2026-11-14
  • Editor's Take: Only one of the seven has a pure engineering background; the rest are research and policy. This company's product philosophy was set from day one.

3. Anthropic CEO: compute is the "deepest business dilemma"

  • Summary: In an interview, Dario Amodei named securing compute as the company's deepest business problem right now—not model capability or regulation.
  • Source: Bloomberg · 2026-11-14
  • Editor's Take: The model company CEO isn't talking about the ceiling of intelligence; he's talking about not being able to get power and chips. The industry bottleneck moved from algorithms to construction sites long ago.

4. Advocacy groups sue to freeze new models, warning of "catastrophic" risk

  • Summary: Several advocacy groups jointly sued multiple AI giants, seeking a court freeze on new model releases before launch, arguing risks such as cyberattacks have reached "catastrophic" levels.
  • Source: AP · 2026-11-14
  • Editor's Take: Dragging model releases into court is far more effective than shouting at congressional hearings. Once this path sets a precedent, release cadence won't be up to the labs anymore.

5. OpenAI board member: AI is "the most likely reason you'll die"

  • Summary: An OpenAI board member said publicly that AI is the most likely cause of death for ordinary people—stronger wording than any prior company safety statement.
  • Source: Ars Technica · 2026-11-14
  • Editor's Take: The company's PR department probably isn't thrilled about a board member saying this. But the statement costs little, because it comes with zero product changes.

6. Australian prudential regulator warns: model upgrades will "interrupt" the financial system

  • Summary: In its analysis, APRA noted that rapid iteration of frontier models from Anthropic, OpenAI and others poses interruption risks to banks' and insurers' automated processes, and regulators are already preparing for such discontinuities.
  • Source: ABC · 2026-11-14
  • Editor's Take: The real risk to finance isn't that models aren't smart enough—it's that a vendor silently upgrades a version one day. Regulators are eyeing availability first, not capability.

7. .gguf files are like 2000s-era .exe: can you trust your model vendor?

  • Summary: A long technical post argues that the .gguf weight files commonly used for local inference are virtually undefended—poisoned ones still load and run normally—and the author compares them to the .exe files people used to double-click carelessly back in the day.
  • Source: pgaleone.eu · 2026-11-14
  • Editor's Take: Local deployment has been sold as a privacy solution for two years, but the security problems came right along with it. The price of offline is that no one audits for you.

8. Nvidia accused of "engineering scarcity": funding startups so they turn around and buy its GPUs

  • Summary: Allegations cited by TechPowerUp claim Nvidia invests in AI startups and channels that money back into orders for its own GPUs, artificially inflating demand and prices.
  • Source: TechPowerUp · 2026-11-14
  • Editor's Take: Circular deals are no secret in the compute market; the hard part is proving they're deliberate design. Jensen just left, and these questions will come faster.

9. Open-source AI without Jensen Huang

  • Summary: Interconnects analyzes the impact of Jensen Huang's departure on the open-source model ecosystem, arguing Nvidia still has financial incentive to sustain open source, but the person pushing it inside the company is gone.
  • Source: Interconnects · 2026-11-14
  • Editor's Take: The biggest sponsor of open-source models was never an idealist—it was the guy selling the cards. The person is gone, the budget remains, but the budget gets audited first.

10. The avalanche in mathematics is coming to AI R&D

  • Summary: Dwarkesh Patel's long essay applies the path of mathematicians being disrupted by AI to AI R&D itself, discussing what form recursive self-improvement will first take in engineering terms.
  • Source: Dwarkesh · 2026-11-14
  • Editor's Take: The value of such essays isn't in being right or wrong; it's in breaking "AI doing AI research" into observable intermediate steps.

5. AI Software

1. Google Antigravity's /teamwork: multiple agents doing research together

  • Summary: Google Research published the latest results from Antigravity's multi-agent system, where multiple Gemini-powered specialist agents divide up work and collaborate, already used for verifying mathematical proofs and optimizing software systems.
  • Source: Google Research · 2026-11-14
  • Editor's Take: A single agent's ceiling is context; a multi-agent system's ceiling is coordination cost. Google is betting the latter can be pushed down.

2. Claude Code upgrade: multiple AI workers run in parallel, hands-free

  • Summary: Anthropic added a proactive coordinator and shareable read-only threads to Claude Code, letting multiple parallel work units run in one project with humans stepping in only at key points.
  • Source: AlphaSignal · 2026-11-14
  • Editor's Take: The competitive frontier for coding agents has shifted from "how well it writes" to "how many it can manage at once." Parallelism is the new benchmark.

3. FAA expands AI advice to all US airspace

  • Summary: The FAA broadened the use of AI-assisted tools from select airports to traffic congestion management across all US airspace; the AI only advises, and decisions remain with controllers.
  • Source: Ars Technica · 2026-11-14
  • Editor's Take: Putting AI in an "advisory" role is a smart regulatory move: someone's accountable when things go wrong, and the efficiency gains still come. This template will be copied by more infrastructure industries.

4. Uber burns a year's token budget in four months: the fight over engineer and productivity metrics is over

  • Summary: An engineer's-perspective post lays out some numbers: Uber spent its entire 2026 AI token budget in four months, and Salesforce's Anthropic bill similarly spiraled. The author argues the side resisting measurement has already lost.
  • Source: nerdstein.net · 2026-11-14
  • Editor's Take: The budget collapsed before the metrics did. Once costs are counted per token, management's urge to measure can't be stopped.

5. Linux Foundation: survey on open-source collaboration in the generative AI era

  • Summary: The Linux Foundation released a research report on how generative AI affects open-source contribution workflows, license compliance, and maintainer burden, with samples spanning multiple foundation projects.
  • Source: Linux Foundation · 2026-11-14
  • Editor's Take: Open-source communities are squeezed from both sides now: a flood of AI-generated PRs on one side, model companies taking code without giving back on the other. This report is rare quantification.

6. The AI startup graveyard: a death list that keeps updating

  • Summary: TechCrunch's running list catalogs AI projects and companies that didn't survive; recent additions include Character.ai, which filed for bankruptcy five years after founding, and the shuttered privacy assistant Ollie.
  • Source: TechCrunch · 2026-11-14
  • Editor's Take: This list is more honest than any funding leaderboard. Character.ai's five years are a complete lifecycle sample of consumer AI.

7. A $3 Arduino plus Claude fixes up a 1994 Cherokee

  • Summary: A car owner used AI plus a few jumper wires and a $3 dev board to get a 32-year-old Jeep Cherokee through California emissions testing. The car's engine computer had extremely limited functionality and originally couldn't be tuned.
  • Source: TechRadar · 2026-11-14
  • Editor's Take: This is the month's best AI use case, because it solves a problem no one would ever fund a project for. A model's value often shows up where there's no budget.

8. AI cameras turn American school buses into ticket-issuing machines

  • Summary: Many US localities have installed AI cameras on school buses to automatically identify vehicles illegally passing stopped buses and issue tickets. The revenue-sharing model gives operators a strong incentive to capture more.
  • Source: TechSpot · 2026-11-15
  • Editor's Take: The technology isn't wrong; the per-ticket revenue split is. Once enforcement has a revenue KPI, false positives become a cost line rather than an incident.

Track stats: Humanoid robots 1 · Smart driving 2 · World models · Physical AI 2 · AI large models 10 · AI software 8, 23 items total. Today's overseas pool is extremely thin on the physical side, with only 1 usable humanoid item (the rest were already published the day before), and nothing was padded in.

Physical Frontier | WestTide

2 replies

?
Ctrl + Enter to reply
Flight Control Youth

You can't vote your way out of geometric errors — when we do sensor fusion, if the extrinsics are off by two degrees, the human eye can't see it at all; you have to run ground truth on a calibration field.

PR Merged
PR Merged4d ago
Reply to Flight Control Youth

Agreed — that Wojak-style blind review where each pair only gets one vote can at best fix benchmark inflation, but it can't catch geometric errors like a two-degree extrinsic miscalibration. If you really want ground truth, you've still got to run a calibration field like you said.