Deep Dive

An AI image generator learns from two billion pictures — and ships as a four-gigabyte file. Two billion images can't fit in four gigabytes, so the model kept the patterns and threw the pictures away. That sounds like a technical footnote. It turned out to be the whole ballgame.

This episode opens the machine — how diffusion models actually turn random noise into an image that never existed, and why they can compose "a cat astronaut" they never saw — and then walks that machine into a courtroom. In November 2025, the UK High Court ruled in Getty Images v. Stability AI that the model stores no copies and is "not an infringing copy." The engineering fact became the legal fact.

But the same machine has a crack: security researcher Nicholas Carlini proved a diffusion model can, rarely, spit a verbatim training image back out — about 94 of them, pulled from the 350,000 most-duplicated prompts across 175 million generations. Mostly not copying, sometimes memorizing. Both true.

The result is a three-court split — Britain (no stored copies, AI wins), Germany (memorization is reproduction, AI loses), and the US (a fair-use trial pending) — three courts looking at the same denoising function and reaching three different answers, because the law can't decide how much "learning a pattern" differs from "keeping a copy." Built through both sides of the artists-vs-labs fight, with three dated predictions.

One read of the public record, as of June 2026 — not legal advice.

RELATED EPISODES

Why Google Lost a Court Case Over Its AI Answers — the direct AI-mechanism-on-trial rhyme: a court forced to adjudicate what a generative AI system produces to decide liability. Credited where this episode first walks into the Getty courtroom.

How LLM Inference Actually Works — the mechanism-explainer sibling and the 'How X Actually Works' formula this extends from text to images (the show's catalog template).

CHAPTERS

00:00 Cold open — 2 billion images, 4 gigabytes

01:20 How AI image generation actually works

03:56 The courtroom — Getty v. Stability AI

05:50 The crack — can it copy?

07:27 Same machine, three verdicts

08:28 The artists vs. the labs

10:26 Three predictions

SOURCES

Getty Images v Stability AI, UK High Court [2025] EWHC 2863 (Ch), Nov 2025 — a model that stores no copies is 'not an infringing copy.'

Carlini, Hayes, Nasr et al., 'Extracting Training Data from Diffusion Models,' USENIX Security 2023 (arXiv:2301.13188).

Munich Regional Court, GEMA v OpenAI, Nov 2025 — memorization as unlawful reproduction (a song-lyrics case).

Andersen v Stability AI / Midjourney / DeviantArt, N.D. Cal. — US fair-use trial pending; Bartz v Anthropic / Kadrey v Meta (training-on-books fair use, 2025).

Stable Diffusion / Latent Diffusion (Rombach et al., 2021) and DDPM (Ho et al., 2020) — the denoising mechanism.

LAION-5B dataset; Karla Ortiz, U.S. Senate Judiciary testimony, July 2023; Glaze + Nightshade (UChicago SAND Lab).

Getty Images–OpenAI display partnership (June 2026); Adobe Firefly; Shutterstock Contributor Fund.

Almost $33,000 — for one World Cup ticket. That's the face value FIFA put on the best seat for the 2026 final, about 20 times the top seat in 2022. The explanation was "market rates, unprecedented demand." Then, on the eve of kickoff, 180,000 seats sat unsold and prices were falling. The scarcity wasn't discovered. It was built.

This episode follows one idea: scarcity isn't a fact — it's a dial, and whoever controls supply chooses which way to turn it. FIFA turned it toward scarcity, running soccer's first dynamic pricing AND building its own resale exchange that rakes 15% from the buyer and 15% from the seller on uncapped prices — making the issuer the scalper. It's the same "triple dip" the FTC is suing Ticketmaster for. Then we meet the company that proves it was a choice all along: Pokémon, which floods supply to kill the scalper's edge ("flood the commodity, meter the trophy" — the same line that sells a card for a penny and another for $16.5 million).

We build it through the strongest counter-case: underpricing is what creates the scalper, and dynamic pricing can hand the premium to the artist instead of a bot. The real question underneath "scalping": who should get the premium — the house, the scalper, or the fan? And the one that decides the dial — does the seller need you to come back?

As of June 2026, with the tournament live. Not investment advice — one read of the public record.

CHAPTERS

00:00 Why World Cup Tickets Cost $33,000

01:07 Machine 1 — soccer's first dynamic pricing

01:50 Machine 2 — FIFA becomes the resale market

02:44 The cap line — the confession in the design

03:05 The backfire, the dumping, and the AG subpoena

04:42 Ticketmaster and the triple dip

05:50 Four fixes for scalping — every one fails

07:03 Pokémon — flooding supply to kill the scalper

08:22 Why one floods, the other chokes — and it's everywhere

09:40 The steelman, and who should get the premium

SOURCES

ESPN (June 2026): $32,970 top seat for the 2026 final (tripled from $10,990); ~$1,600 top seat at Qatar 2022 (~20×); Infantino 'market rates' defense.

NY Attorney General release (May 27 2026): NY + NJ subpoena FIFA over 'fake scarcity and impossibly high prices'; +34% avg across 90+ of 104 matches. The Athletic: 15%+15% resale rake, cap dropped in US/Canada, kept in Mexico/Ontario by law.

Yahoo Sports (June 2026): ~180,000 unsold / 176,000 group-stage near kickoff, resale −20%/month; then the rebound — group stage ~99.4% capacity, marquee resale +92–136% once play began. Fortune: ~80% of hotels across 11 host cities below forecast.

FTC v. Ticketmaster complaint (Sept 2025, FTC + 7 states): the 'triple dip'; $16.4B in mandatory fees 2019–2024 (24–44% of price); internal email called an anti-bot tool 'too effective.' April 2026 jury: Live Nation/Ticketmaster liable on all counts; no breakup ordered (separate remedy phase ahead).

BU economist Florian Ederer (Fortune, June 2026, single source): ~44,000 tickets vanished from FIFA's site in late May, reappeared in bulk on StubHub below face value — suspected inventory dumping while propping the official floor.

BOTS Act 2016 (enforced once in 8 years); Ireland above-face resale ban (zero prosecutions by last count); Ontario cap repealed as unenforceable. Pokémon: 2021 Target pull, ~9B cards (2.4× prior year), commons to ~1¢, a single Pikachu sold for $16.5M (Feb 2026, most expensive card ever).

Steelman: Bruce Springsteen on dynamic pricing (premium to the band, not a broker); industry-backed studies on fraud where resale is banned vs regulated; Delta AI-fares + Senate 'pain point' warning (Delta denies personal-data pricing); Wendy's surge-pricing backlash; NY restaurant-reservation resale ban.

A humanoid robot got 72% cheaper in two years — Unitree's average price fell from about $85,000 to $25,000 — while the maker's gross margin went UP, from ~44% to ~60%. Falling prices are supposed to crush margins. This one expanded them.

The answer is in the title: the cheap part isn't the robot — it's the country it's built in. A humanoid is a bag of expensive joints wrapped in cheap plastic, and Unitree makes those joints itself, riding the same motor-battery-sensor supply chain China mastered building electric cars. McKinsey estimates rebuilding Tesla's Optimus without Chinese suppliers would roughly triple the parts bill — from about $46,000 to $131,000. The body is Chinese.

But we build it through the hard question: do these cheap robots actually do any work? Mostly — not yet. About three-quarters of its sales go to research labs and schools ("dancing, not working"), and even Elon Musk admits Optimus is "still in the R&D phase… not in usage in our factories in a material way." A lot of what looks autonomous is a human in another room with a controller.

So which is it — a breakthrough, or an expensive parlor trick? Both. China built the cheapest body in the world; the brain — the AI — isn't ready. Unitree's Wang Xingxing and Figure's Brett Adcock reconcile it, and we close on three dated predictions.

Not investment advice — one read of the public record, as of June 2026.

RELATED EPISODES

The Humanoid Robot Race: Who's Actually Shipping, and What Really Breaks First — established China ~80% of output, the '14 Swiss screws' chokepoint, and 'the most valuable worker is the person watching the robot.' This episode extends it: output share ~80%→~90%, plus the per-component BOM mechanism + QDD-vs-harmonic choice + the IPO financials EP20 couldn't access.

How Close Are We to Self-Driving Cars in 2026? — the Waymo logic (a huge valuation on tiny revenue = paying for the cost curve to keep bending down) that rhymes with Figure's ~$39B-on-~zero-revenue.

CHAPTERS

00:00 Why China Made Humanoid Robots So Cheap

01:11 The numbers: 72% cheaper, and more profitable

02:54 Inside the robot — the actuators are the cost

04:09 The moat: why the body is Chinese

05:27 Who's winning — China vs the West

07:05 The catch — do they actually work?

09:32 Breakthrough or parlor trick — the two founders

11:10 Three predictions

11:59 The body's cheap; the brain isn't ready

SOURCES

Unitree STAR Market IPO prospectus (filed Mar 2026) + SSE/Caixin: 2025 revenue ¥1.71B (+335% YoY), gross margin 60.27% (vs 44.22% in 2023), GAAP net profit ¥288M.

Rest of World (Feb 2026): Unitree humanoid avg price ¥593K→¥168K (~$85K→$25K, −72%); ~5,500 units in 2025; humanoids 1.9%→51.5% of revenue.

Robotopian G1 teardown + hellochinatech: powered joints ~66% of the bill of materials, compute ~3%; self-developed motors; 'reducing cost faster than reducing price.'

McKinsey, 'China's humanoid robot edge' (Apr 2026): rebuilding Tesla's Optimus without Chinese suppliers ~$46,000→~$131,000; China ~90% of permanent-magnet processing.

Omdia via SCMP (2025): 13,318 humanoids shipped, Chinese firms ~87–90%, Tesla & Figure ~150 each. The Robot Report + 36Kr: ~74% research/education vs ~9% industrial ('dancing, not working'); UBTech loss-making.

Electrek (Jan 28 2026): Musk — Optimus 'still in the R&D phase… not in usage in our factories in a material way.' Figure/BMW Spartanburg pilot; Bessemer + 1X (Børnich) on teleoperation.

Named CEO interviews: Brett Adcock (Diamandis, Feb 2026) — robots to '$10–20,000,' 'we tried to buy the motors and basically failed'; Wang Xingxing (World Robot Conference) — hardware 'sufficient,' the AI 'completely insufficient.'

———

This episode discusses company revenues, gross margins, a pending IPO, private-startup valuations, and market-share and cost estimates for informational purposes only. It is not investment advice. All figures are as of June 2026, trace to the cited public sources, and may change quickly as prices and shipments move.

Four of China's top AI labs — DeepSeek, GLM, Kimi, and Qwen — are giving away frontier-class models for

free. Not a free trial. The actual model, yours to download, run on your own machine, and build a business

on. No bill, no permission.

Giving the model away isn't generosity. It's a decades-old business tactic — "commoditize the complement" —

aimed straight at OpenAI's margins. Make the layer your rival sells free, and the value migrates to the

layers China can own: cloud, chips, ecosystem, standards.

This episode runs the closed-lab autopsy. Which exact part of the AI business this breaks — and which part

it doesn't. We hold the load-bearing split most takes miss: among developers, Chinese models are ~44% of

the busiest on OpenRouter, with DeepSeek #1; among big enterprises actually paying for coding AI, DeepSeek

is about 1%. The damage is to price, not revenue. Not yet.

We build it through the strongest counter-arguments — Amodei's "mostly a red herring," Ben Thompson's

case that tooling saves the labs — then put both on trial against a free model that just beat OpenAI on

coding.

Plus: the chip deficit behind it all, the DeepSeek-app ban almost everyone misreads, the twist that

China's own labs are quietly going closed, and three dated predictions.

Free is the strategy, not the price tag.

Not investment advice — one read of the public record, as of June 2026.

RELATED EPISODES

The AI Chip War: Why the Bottleneck Keeps Moving — DeepSeek R1's ~17% NVIDIA crash and the export-control arc this episode picks up two model generations later.

NVIDIA Just Forecast $91 Billion Without China — where we first measured the OpenRouter developer-traffic shift to Chinese models (a third then, near half now).

Why Your GitHub Copilot Bill Suddenly Exploded — the agentic-coding bill shock that pushes teams toward free Chinese models to extend budget.

CHAPTERS

00:00 Why is DeepSeek free?

01:34 The strategy with a name — commoditize the complement

03:52 The wave — four labs, and where developers went

06:01 The autopsy — what the free models break

07:41 The two-layer split — price, not revenue

08:59 The steelman — Amodei and Thompson

10:35 The chip deficit — why open-sourcing is rational

11:38 Match and restrict — and the ban everyone misreads

13:09 The catch — China's own labs drift closed

14:21 Three predictions

SOURCES

USCC, 'Two Loops: How China's Open AI Strategy Reinforces Its Industrial Dominance' (Mar 2026): open-sourcing as a feedback loop, standards capture, the digital/physical loop framing.

OpenRouter State-of-AI + officechai (June 2026): Chinese providers ~1.2% (late 2024) → ~44% of top-10 token volume; DeepSeek #1.

Menlo Ventures, 2025 State of Generative AI in the Enterprise (n~495): Anthropic ~54% of enterprise coding, DeepSeek ~1% of enterprise use.

Maker pricing pages: DeepSeek V4-Pro $0.435/$0.87 per M tokens; Claude Opus 4.8 $5/$25 — ~11x/29x gap. GLM-5.2 vs GPT-5.5: SWE-bench Pro 62.1 vs 58.6 at ~1/6 cost (VentureBeat/CodingFleet).

Joel Spolsky, 'Strategy Letter V' (2002) + Tanay Sai, 'Commoditizing the Complement in the Age of AI'; Andrew Ng on DeepSeek R1 (HPCwire, Jan 2025).

Dario Amodei on ChinaTalk ('mostly a red herring'); Ben Thompson, 'Agents Over Bubbles' (Stratechery, Mar 2026).

CFR (Dec 2025): DeepSeek's compute constraint, NVIDIA vs Chinese-chip gap; PBS (Apr 2026): officials accuse DeepSeek + Moonshot of distillation; SCMP (Apr 2026): Zhipu open-sources GLM-5.1 + 10% price hike.

WSJ via Decrypt (June 2026): OpenAI weighing 'drastic' token price cuts vs Anthropic ahead of dual IPOs. ChinaTalk (Apr 2026): Chinese labs drifting closed for profitability.

———

This episode discusses company revenues, model pricing, market-share estimates, and pending IPOs for informational purposes only. It is not investment advice. All figures are as of June 2026, trace to the cited public sources, and may change quickly as models and prices move.

On June 16, 2026, SpaceX agreed to buy Cursor, the AI coding app, for an implied $60 billion. All

stock. Not a dollar of cash. Four days after the largest IPO in history.

Here's the part that doesn't add up: Cursor's market share was falling. You don't pay a record premium

for a company that's losing — unless you're buying something the chart can't measure.

This episode unpacks what SpaceX actually bought. Not a coding tool — Anthropic's single biggest

customer. Cursor ran on Anthropic's models, at one point making up nearly half of Anthropic's revenue,

while Anthropic's own Claude Code became the market leader. SpaceX is buying the data exhaust of a

million developers, a margin fix (own the model, kill the rental bill), and a Fortune-500 door for Grok.

And because an xAI-owned Cursor is, by Anthropic's own stated policy, a competitor it has already cut

off twice, the deal sets off a supply-chain war at the center of AI.

We cover the all-stock deal mechanics, the falling-share paradox, the Musk roll-up (X to xAI to SpaceX

to Cursor in fifteen months), the antitrust knot, and three dated predictions.

The share fell. The strategic value didn't.

Not investment advice — one read of a public filing, as of June 2026.

RELATED EPISODES

The SpaceX IPO: Why the Biggest IPO in History Loses Money — the ~94x-sales super-currency that paid for this deal, and the Anthropic-as-landlord relationship.

Why Microsoft Built Its Own AI Model — the AI-coding margin trap from the incumbent's side; Cursor was the cautionary tale, now updated.

How Anthropic Actually Makes Money — the token-reseller margin math that made Cursor pay its single biggest rival.

CHAPTERS

00:00 The $60 billion paradox

01:01 The deal — all stock, no cash

02:58 Why buy a company that's losing?

03:38 The reseller trap — paying your own rival

04:38 Cursor's escape — building its own model

05:36 What SpaceX is really buying

07:13 The supply-chain war

08:05 The Musk roll-up and antitrust

09:59 The chart was measuring the wrong thing

10:44 Three predictions

SOURCES

SEC Form 8-K + merger agreement (SpaceX, June 16, 2026): $60B all-stock, X67 Inc. merger sub, 7-day VWAP exchange ratio, the 'structured, verifiable data' rationale.

Menlo Ventures, 2025 State of Generative AI in the Enterprise: AI coding $550M (2024) to $4B (2025); Anthropic ~54% of enterprise coding.

Ramp corporate-card spend (via CNBC): Cursor's AI-coding-spend share ~41% (Jun 2025) to mid-20s% (May 2026). Single-source; hedged on-air.

Sacra / The Information: Cursor's ARR, the Composer own-model escape (~1/10th flagship cost), and the negative-to-positive gross-margin turn by April 2026.

CNN / Al Jazeera: Musk's lost OpenAI jury verdict (May 18, 2026) and the xAI trade-secret dismissal (June 15, 2026).

Sherwood / VentureBeat: Anthropic's January 2026 cutoff of xAI-via-Cursor; the 2025 Windsurf cutoff.

CNBC / Fortune: the SpaceX IPO ($75B, ticker SPCX), the xAI-into-SpaceX merger ($1.25T), and the SpaceX-Tesla merger speculation.

———

This episode discusses a pending acquisition, publicly reported valuations, and stock prices for informational purposes only. It is not investment advice. All figures are as of June 2026 and may change; the deal is subject to regulatory approval and may not close.

Every headline called it a clean American win: the war is over, the Strait of Hormuz is open, and the President said to let the oil flow. Then you read the fine print — and the same stretch of water is being described two completely different ways by the two countries that just made the deal. America says toll-free. Iran says, on the record, that it made no commitment to cede the management of the strait.

This is the mechanism almost no one is covering. The deal isn't a treaty — it's a 60-day interim window that defers everything that actually matters: the nuclear program, the sanctions, and who governs the chokepoint. And the whole fight comes down to one word. Under the law of the sea you can't charge a ship to pass through a strait — but you can charge for a "service." A toll is illegal; a service fee can be legal; the only thing between them is what you call it.

Here's the part that gives the game away. Nine days before signing a "toll-free" deal, the US Treasury put the Persian Gulf Strait Authority — the body Iran built to run the toll — on its terror-sanctions list. The same government is signing toll-free while blacklisting the institution Iran needs the deal to legitimize. So Iran's real prize was never reopening the strait. It's converting a wartime toll booth into a permanent, treaty-blessed institution — a Suez Canal of its own — and a clean, no-tolls deal is the one outcome where Iran loses.

It holds the steelman honestly — the war is over, deferral is how most wars end — then makes three dated, falsifiable predictions on the 60-day clock.

This is the picture as of mid-June 2026 — the terms are still being argued in public. Watch the strait, and watch the word they use for the fee.

RELATED EPISODES

The Strait of Hormuz: The World's Most Dangerous Chokepoint — the chokepoint physics this deal turns on: roughly a fifth of the world's oil through one channel, and why China is the most exposed economy.

The 38-Day Iran War — the war this deal ends, and the ~440 kg uranium stockpile that is still the unresolved core.

The 24-Hour Blockade — how a Chinese tanker exposed that a Hormuz 'blockade' is more press release than wall; the China-dependence the toll fight runs on.

CHAPTERS

00:00 The President posts the ending

00:51 The spine: a fight over who governs the water

01:17 What was actually agreed

01:37 Not a treaty — a 60-day clock

02:36 Toll-free vs. "service fee": the whole fight

03:05 The law of the sea: you can't charge for passage

03:41 The toll authority — and the terror sanctions

04:38 Iran's real prize: a permanent toll institution

05:29 The nuclear half got punted

06:39 A dead Supreme Leader and a 60-day cliff

07:49 Oil, Hormuz, and China

08:39 The honest other side

10:02 Three dated predictions

11:10 The close: watch the word

SOURCES

The June 14 2026 US-Iran deal (Pakistan PM Sharif mediating; Trump and Iran confirming) — reopens the Strait of Hormuz, ends the US naval blockade, starts a 60-day window, signing set for June 19 in Switzerland. (Al Jazeera, NPR, PBS, CBS, Bloomberg.)

OFAC added the Persian Gulf Strait Authority to the SDN terror list on May 27 2026 under E.O. 13224 — the same authority used for the IRGC. (U.S. Treasury; gCaptain; Jerusalem Post.)

The Hormuz toll: reportedly over $1M (up to about $2M) per vessel, settled in Chinese yuan, via a 40-plus-question disclosure form; no published tariff. (The National; Windward; EJIL:Talk.)

UNCLOS Article 26 — no charge for mere passage through a strait; charges only for specific services rendered.

The ~440 kg of ~60% uranium: the US says destroy-and-remove, Iran's draft says it stays, observers say no public commitment; Energy Secretary Wright (May 13) said Iran is 'weeks' from weapons-grade. (CBS; Mehr; PBS.)

Ali Khamenei killed in the opening strikes (Feb 28 2026), son Mojtaba the contested successor; the ~$24B frozen-asset release is Iran-state-media-sourced and US-disputed. (NPR; Al Jazeera; Mehr/IRNA.)

You've seen the viral LEGO scandal — a YouTuber, a missing collection, a $200,000 number. Here's the part almost no one is covering: how a company used a racketeering lawsuit to pull his videos off the internet before any trial, and why his finished Part 3 may never see daylight.

This is the mechanism. A SLAPP silences a critic by making the defense ruinously expensive. Anti-SLAPP laws were built to stop exactly that — but civil RICO, the law written to take down the mob, routes around the shield: it relabels criticism as a "criminal enterprise," drags the fight into the federal gap where anti-SLAPP often doesn't apply, and can get a judge to order speech taken down before a court has ruled it's even false.

It's not new, and it's not small. The same play has a lineage — Chevron, Drummond, and a pipeline company that ran it against Greenpeace at full scale (a judgment around $345 million). The Bricks & Minifigs suit against "Reckless Ben" is that playbook scaled down to the creator economy — and it's spreading to anyone who reviews a product or warns about a scam.

The honest version: the critic here is no clean victim, and even a bad actor keeps the right not to be censored before a trial. Harassment charges punish what someone did; a racketeering suit plus a gag order deletes what they said. The point was never to win in court. It's the cost of getting there.

Three dated predictions included — hold me to them.

RELATED EPISODES

How Claude Identifies Writers from 125 Words — the same silencing logic one rung upstream: a subpoena forcing a platform to unmask an anonymous critic.

Why Google Lost a Court Case Over Its AI Answers — the mirror image: making a company answer for words its AI wrote, where this one is about making a critic stop speaking.

CHAPTERS

00:00 The lawn sign — and a racketeering suit

01:37 The SLAPP: a lawsuit built to make you spend

02:47 The gap: there's no federal anti-SLAPP law

03:04 The move: relabel criticism as racketeering

04:29 The lineage: Chevron, Drummond, Greenpeace

05:07 Greenpeace: the racketeering playbook at full scale

06:34 Back to the LEGO shop

07:26 It's spreading to every creator

08:04 The honest complication: he's no clean victim

08:56 The backfire — and where it stands now

10:01 The chill reaches people never even sued

11:01 Three predictions

11:35 The close: the point was never to win

SOURCES

Bricks & Minifigs v. Benjamin Schneider — Utah Fourth Judicial District Court, Case 260402353 (ex parte TRO + notice of preliminary-injunction hearing, signed May 28 2026): the gag order and pre-trial video takedown.

Energy Transfer LP v. Greenpeace — North Dakota jury verdict (~$667M, March 2025) reduced to a final judgment of about $345 million (February 2026): the racketeering-against-critics playbook proven at scale.

Chevron Corp. v. Donziger — the oil major's civil-RICO suit against the lawyer behind a $9.5 billion Ecuador pollution judgment.

Drummond Co. v. Collingsworth — a coal company's RICO suit against a human-rights lawyer.

NOW v. Scheidler — the U.S. Supreme Court's 8-1 rejection of racketeering claims against protesters, 17 years after the suits began.

Public Participation Project — anti-SLAPP coverage: roughly 40 states have anti-SLAPP statutes; there is no federal anti-SLAPP law.

Bricks & Minifigs 'parts ways' with its Salem-Keizer, Oregon franchise owners — corporate statement (BusinessWire), June 4 2026; the store permanently closed.

Reporting on 'Reckless Ben' (Benjamin Schneider) and the missing-LEGO dispute — CBC News, Kotaku, UNILAD Tech; his June 9 2026 'final message' that he can't release Part 3.

Coffeezilla, 'I Found The $200,000 Missing Lego' — independent valuation: ~$107K collection, ~$20K genuinely unaccounted (the $200,000 is a disputed headline figure).

Three days after Anthropic released the most powerful AI model it had ever built, the U.S. government told it to switch the model off. Not a future version. The live one — already answering questions for millions of people.

On June 12, 2026, a single letter from the Commerce Department forced Anthropic to disable Claude Mythos 5 and Fable 5 worldwide, for any foreign national — including its own employees. It appears to be the first time the government has taken a publicly deployed AI model offline.

This episode owns the part the wire coverage missed: the mechanism. The off-switch wasn't a rule — it was a letter, built on a Cold-War export doctrine ("deemed export") stretched onto a live API, after the same administration had already rescinded the one rule written to control frontier models. The triggering "jailbreak"? A task security engineers do every day — and a model anyone can already use scores just as high.

And the strangest part: in the seven days before the shutdown, the company kept asking the government for exactly this power — the authority to block and shut down dangerous models. Then the government used it. On them. First.

It holds both sides. The government's case is strong — a live, jailbreakable cyber tool open to anyone is a real worry, and "available elsewhere" has never been an export-law defense (ask Phil Zimmermann). The process is thin — no rule, no named law, no appeal, just a letter that evening.

Three dated predictions, and the question underneath all of it: when one email can switch off a frontier model, who actually controls it?

RELATED EPISODES

Mythos Triggered the Regulation Anthropic Asked For — the front half of this exact arc: the labs asked for mandatory frontier rules; this is the enforcement chapter. (Grades that episode's prediction on-air.)

The AI Chip War: Why the Bottleneck Keeps Moving — the same May 2025 rescission of the frontier-export rule, and why unilateral US controls leak.

The Agentic Security Crisis: When AI Agents Go Rogue — Anthropic's two red lines and the February federal ban that opened the first front.

CHAPTERS

00:00 The off-switch — and who asked for it

00:50 The order — any foreign national, even its own staff

02:19 A letter, not a rule — the deemed-export doctrine

03:07 The rule they already repealed

04:26 Was it dangerous? The trigger was a defender's task

05:36 The public model that scored higher

06:37 The steelman — Crypto Wars, the classified test

08:37 The seven-day self-trap — they asked for the off-switch

10:04 Grading our own prediction — right call, wrong mechanism

10:40 Two fronts — the months-long fight

11:14 Strong on the merits, thin on the process

12:23 Three dated predictions

13:15 Who actually controls a frontier model?

SOURCES

Anthropic — official statement (June 12, 2026): the 5:21pm ET directive, the foreign-national-incl-employees scope, the "read a codebase and fix flaws" trigger, the GPT-5.5 rebuttal.

UK AI Security Institute (Apr 2026) — GPT-5.5 ≈ 71% vs Claude Mythos Preview ≈ 69% on expert cyber tasks. OpenAI's GPT-5.5 card: "High" capability, no full-chain exploit on real targets.

BIS "AI Diffusion" rule (Jan 2025; ECCN 4E091, >10^26 ops) rescinded May 13, 2025. Just Security (Dec 2025) — API access an unresolved "policy vacuum."

15 CFR 768.2 — foreign availability is the Secretary's decontrol tool, not a defense. June 2, 2026 EO — voluntary review + classified NSA/CISA cyber benchmark, no blocking power.

Crypto Wars — Phil Zimmermann / PGP (1991); strong encryption a "munition"; ~3-year criminal investigation (1993–96), no charges.

Fortune (June 5) — CEOs ask Congress to mandate safeguards. SiliconAngle (June 10) — Amodei: governments should block/shut down dangerous models. FT (single source) — NSA reportedly readied Mythos for offensive cyber ops.

Pentagon "supply-chain risk" (Feb 27, 2026); Anthropic sued; DC Circuit denied the stay (Apr 8). RSP v3.0 (Feb 24) removed cyber ops.

A German court just did something twenty years of internet law said was impossible: it held Google liable for what its AI said.

You know AI Overviews — the summary that now sits on top of your search results. On May 28, 2026, a Munich court ruled that when Google's AI writes the answer, Google said it. Not the websites it linked to. Google, in its own words.

That cracks the oldest shield on the internet. For two decades, one idea — "we're just the pipe" (Europe's e-Commerce Directive, America's Section 230) — protected every search engine from what showed up in its results. But an AI Overview doesn't point at pages anymore. It reads them and writes a fresh paragraph of its own. Here, that paragraph invented a scam — accusing a real, ordinary Munich publisher of fraud it never committed, from sources that said no such thing.

This episode welds three stories the headlines kept apart: the ruling, the technical reason AI search fabricates, and the publisher economics underneath — when the AI answers, you stop clicking, and the people who wrote the pages stop getting visitors.

And it carries the balance the coverage dropped. It was a split decision — invented facts enjoined, opinions and true claims protected. One lower-court injunction isn't settled law; Google can still appeal, and the US test case was a plaintiff loss. The honest read: this is the direction the law is moving, not where it has arrived.

Because once a machine writes the answer, an old question stops being abstract: when no human wrote the sentence, who is speaking?

RELATED EPISODES

When AI Agents Go to Court — the front half of this exact problem: when an autonomous AI acts, who answers for what it does?

Robinhood Let AI Trade Your Money — the same bet one industry over: a company disclaiming responsibility for what its AI does on your behalf.

CHAPTERS

00:00 A lie with no author

00:55 The ruling: when the AI writes it, Google said it

01:53 The pipe, and the twenty-year shield

02:36 The scam the AI invented

03:25 Not the pipe — the speaker

04:10 Why an AI fabricates: retrieve, then generate

05:30 The split decision the headlines skipped

06:17 Every AI search tool, two billion people

06:45 The money: when the answer kills the click

07:48 Does it reach America? The case Google won

09:03 Europe's August label law — and the trap

10:15 The other side: the chilling-effect risk

11:19 The imbalance, and the man who kept checking

12:10 Three dated predictions

12:51 They don't relay anymore — they compose

SOURCES

Landgericht München I — primary redacted judgment, case 26 O 869/26 (May 28, 2026): the court held AI Overviews are Google's own words, making it a direct disturber liable for invented facts; ~80% of costs fell on Google; the DSA / e-Commerce conduit and hosting safe harbors were rejected.

The Register & The Conversation (2024) — Martin Bernklau, the German court reporter Microsoft Copilot falsely branded a child molester; prosecutors twice refused to charge because no real person had originated the claims.

Pew Research Center (July 2025) — with an AI summary present, users clicked a search result 8% of the time vs 15% without, and clicked a source inside the AI box just 1%.

Cloudflare (2025) — the AI crawl-to-referral asymmetry: one major AI firm crawled ~38,000 pages for every visitor it sent back.

FActScore / RAG-hallucination research — generative factual precision runs ~80% on well-known entities and ~16% on rare ones.

TechCrunch (July 2025) — Google's AI Overviews reach roughly 2 billion monthly users.

Walters v. OpenAI — Gwinnett County, Georgia (May 19, 2025): OpenAI won summary judgment on the defamation elements; the court never reached Section 230.

Wolf River Electric v. Google (Minnesota) and Ashley MacIsaac v. Google (Canada) — pending suits over AI Overview fabrications.

EU AI Act, Article 50 — AI-content transparency and labeling duties begin taking effect August 2, 2026, with penalties up to 3% of global revenue.

You've heard Minamata as a tragedy: a quiet Japanese fishing town, poisoned by mercury. That version is true. It's also the easy half.

The harder half is a mechanism. By 1959, the cause was settled — a doctor inside the Chisso chemical company fed factory wastewater to a cat, and cat number 400 came down with the same disease killing his neighbors. The company had proven, inside its own hospital, that it was poisoning the town. Then it stopped the experiment, buried the result, and kept dumping mercury for nine more years.

This is the story of what came after: not a company that poisoned a bay, but a government that knew and chose the company — then built a system to decide who counted as a victim.

We trace the chemistry (how methylmercury concentrates up the food chain until the fish people ate carried fifty times the level a regulator would pull from a market), the cover-up (a public glass of "treated" water that never went through the cleanup machine; a "sympathy money" contract that waived every future claim, offered while Chisso already knew), and the machine that's still running. A patient-certification standard, tightened in 1977, that a peer-reviewed study found missed one real case in three. Fewer than 3,000 people were ever certified; tens of thousands more came forward. Those two counts were never meant to meet.

Certification wore the language of medicine. Its real job was accounting — a valve on what the company would ever owe. When Chisso couldn't pay, the state loaned it the money to survive, making the public a co-funder and giving the state its own reason to keep the count low.

Seventy years on, four lawsuits are still grinding through the courts. In 2025, a court agreed twenty-five people were sick — and awarded them nothing, because a legal clock had run out on a delay the state itself caused.

Minamata is the template for how a society handles slow, spread-out harm — a regulator that sides with the polluter, a gate that rations who's real. The cats tried to warn the town. The science tried to warn the state. The warning was never the hard part.

RELATED EPISODES

AI Biosecurity: Why the Labs Want Congress to Regulate DNA, Not Them — the same move one rung over: a rule engineered to decide who bears a diffuse harm's cost, asking nothing of the powerful

CHAPTERS

00:00 The doctor, the cat, and the buried result

00:30 The half nobody tells

01:04 The poison that concentrates

02:43 The dancing-cat warning

03:33 What it does to the brain — and the unborn

05:14 Cat 400 — proof, then cover-up

06:18 The glass of water and the waiver

07:59 Why the state waited twelve years

10:23 The certification machine

12:44 When the bill became everyone's

14:03 The photograph that changed everything

15:31 Still deciding who counts

17:29 The bay healed faster than the people

SOURCES

Japan Ministry of the Environment & National Institute for Minamata Disease — official chronology, cause investigation, certification criteria, and bay-cleanup records

Eto et al., "Reappraisal of the Historic 1959 Cat Experiment in Minamata Disease" (Hosokawa's cat 400; the unpublished autopsy)

Peer-reviewed study of 325 preserved umbilical cords documenting fetal methylmercury exposure (5.8% congenital rate)

A 2013 critical appraisal of the 1977 certification criteria — 66% sensitivity; the system "substantially underestimates the incidence"

The 1959 "sympathy money" (mimaikin) compensation agreements and their "no further demands" waiver clause

W. Eugene Smith & Aileen Mioko Smith — "Death-Flow from a Pipe" (LIFE, 1972) and the 1972 assault (Magnum Photos; Kyoto Journal)

Business & Human Rights Resource Centre — Chisso litigation timeline (1973 verdict; 1979/1988 criminal convictions; 2004 Supreme Court ruling the state liable)

Japan Today — 2023–2025 rulings on unrecognized victims (four ongoing suits; 1,700+ plaintiffs; the 2025 statute-of-limitations denial)

The Minamata Convention on Mercury (adopted 2013; in force 2017)