AI ACTUALLY
Issue No. 18 — Sunday, June 14, 2026
On Tuesday, Anthropic handed the public the most capable AI model it had ever released. On Friday at 5:21 p.m., the US government told it to switch the thing off — and within hours, it was gone for everyone on Earth. That, plus a rocket company becoming worth $2 trillion, Jeff Bezos quietly spending $41 billion to automate engineers, and a pocket translator straight out of science fiction. Coffee first. Then this.
The most powerful AI the public ever got lasted three days
On Tuesday, Anthropic released Claude Fable 5, which it billed as the most capable model it had ever made generally available — a new tier it calls “Mythos-class,” strong at software, science, and finding flaws in code. Alongside it came Mythos 5: the same model with a few guardrails loosened for vetted cybersecurity and biology users.
Then, per Anthropic’s own statement, a government export-control order arrived Friday at 5:21 p.m. Eastern. Citing “national security,” it barred any foreign national anywhere — including Anthropic’s own foreign employees — from using either model. There was no clean way to enforce that, so Anthropic did the one thing that guaranteed compliance: it disabled both models for everyone, worldwide, within hours. Every other Claude model kept running.
The trigger, as far as Anthropic can tell, is a “jailbreak” — a trick for getting Fable to slip its own guardrails and hunt for software vulnerabilities. Anthropic says it watched a demo, found the technique surfaced a handful of already-known, minor flaws, and that other public models can do the same. It’s complying, but it disagrees, warning that if this became the standard it would “essentially halt all new model deployments” across the industry. It also apologized and called the whole thing a misunderstanding. A model hundreds of millions of people could use Friday morning was gone by Friday night — and the company that built it didn’t make the call.
Why it matters. The interesting part isn’t the jailbreak. It’s that a private company built something, marketed it as powerfully dangerous, and the government took it at its word and reached for an off switch nobody knew existed. For one weekend, the most advanced AI available to the public wasn’t controlled by the company that made it or the people paying for it. It was controlled by a letter.
Anthropic’s CEO wants to be regulated. Harder.
Days before the shutdown, Anthropic CEO Dario Amodei published an essay arguing that AI now moves faster than anyone governing it. He opened by comparing Washington to Treebeard, the Lord of the Rings tree so slow that saying hello takes all afternoon. His fix is an odd thing to hear from someone running a market leader: regulate me, and do it faster.
The asks include giving regulators the power to “ground” a frontier model the way aviation authorities ground a plane, screening new models across four risk areas before launch, fast-tracking AI-designed drugs, and tightening chip export controls. And — the part worth slowing down for — a plan for what he calls unprecedented unemployment: investment accounts that hand ordinary people shares in the AI companies, plus a version of universal basic income. Then, 48 hours later, the government grounded one of his models for him. He presumably did not mean it quite that literally.
Why it matters. “Regulate me harder,” from the company with the most to lose, is easy to wave off as theater — and skeptics will. But Amodei has been making this case since before it was convenient, and the timing did him a strange favor: the week he warned these models had become national assets, the government treated one like exactly that. The notion that the public should own a slice of the AI giants now keeps surfacing from every direction at once — we watched Washington float it three weeks ago. When the people building the thing start proposing how to share it, the conversation has moved.
SpaceX had the biggest IPO ever — and it’s selling data centers in space
On Wednesday, SpaceX priced the largest IPO in history: $75 billion raised, at roughly $1.75 trillion. On Friday it started trading, jumped 19% on day one, and closed worth more than $2 trillion — instantly one of the ten most valuable public companies on the planet, with record numbers of small investors piling in.
Two things make this an AI story, not a rocket story. First, SpaceX quietly absorbed Elon Musk’s AI company, xAI, earlier this year, so Grok now lives inside the rocket company. Second — and this sounds invented — SpaceX used the run-up to preview “AI1,” a solar-powered satellite built to run AI chips in orbit. The pitch: data centers on the ground are fighting towns over power and water, so put them in space, where the sun never sets and nobody files a zoning complaint. Each satellite would carry about as much computing muscle as one of Nvidia’s top server racks, and Google and Anthropic have reportedly already signed on as customers.
SpaceX is also just the opening act. It’s the first of a trio of mega-listings: Anthropic filed confidential IPO paperwork two weeks ago, and OpenAI filed its own draft days later, both expected to go public above $1 trillion. But OpenAI is hedging on when. In an internal note, Sam Altman reportedly floated that if “self-improving AI” — systems that improve themselves — starts looking close, it might be smarter to delay the IPO than rush to market mid-breakthrough. The other complication is the bill: OpenAI is weighing compute commitments running into the hundreds of billions, including a reported $500 billion data-center campus in Ohio. SpaceX just proved the money is waiting. OpenAI is deciding whether it wants it yet.
If “renting compute from a rocket company” rings a bell, it should — back in Issue No. 17 we covered Google paying SpaceX $920 million a month for chips because it couldn’t build fast enough. The rocket company is now the landlord for the entire AI boom. One detail the champagne skipped: SpaceX lost $4.3 billion last quarter, and at least one analyst pegs its real worth at about a third of the IPO price. Data centers in space; losses on Earth.
Why it matters. The most valuable thing in AI right now isn’t a model — it’s the power and chips to run one, and increasingly the land to put them on. A rocket company just became one of the biggest on Earth by promising to solve that in orbit, and two more trillion-dollar listings are lining up behind it. Whether or not the satellites ever fly, the market just bet $2 trillion that the real bottleneck is physical.
Bezos spent $41 billion to automate the engineer, not the chatbot
While everyone watched the rockets, Jeff Bezos resurfaced with an AI company of his own. It’s called Prometheus, it just raised $12 billion at a $41 billion valuation, and it is pointedly not a chatbot. Bezos wants to build what he calls an “artificial general engineer” — software that designs and manufactures physical things: jet engines, drug compounds, complex machinery.
His complaint is that physical invention is absurdly slow. Asking for a jet engine with 10% more thrust, he says, can be a ten-year project. Prometheus aims to make the loop from idea to finished product ten times faster. It’s about 150 people, with offices in San Francisco, London, and Zurich, and backers including JPMorgan, Goldman, and BlackRock. Asked the obvious question — does this replace engineers? — Bezos went the other way, predicting AI will create “more than 10x” the opportunities and coining the phrase “labor scarcity.” It’s a sunny forecast from a man whose other company, Amazon, has cut tens of thousands of jobs while leaning into automation.
And Bezos isn’t even alone in the idea. A far smaller startup, P-1 AI, is chasing the same thesis with a character it calls Archie — an AI engineer pitched, very deliberately, as a colleague rather than a tool. By P-1’s own description, Archie is currently about as capable as a junior mechanical and electrical engineer, and you “hire” him roughly the way you’d bring on an outsourcing firm. The company is starting with the least glamorous, most in-demand corner of engineering — data-center cooling and power — before expanding into automotive and aerospace, and it’s backed by names like Google DeepMind’s Jeff Dean and an OpenAI executive. (The name is an homage to a sentient AI from a 1977 sci-fi novel, which tells you something about the ambition.) The gap is the punchline: Prometheus is worth $41 billion; P-1 is working off a seed round. Same bet, four orders of magnitude apart.
Why it matters. Most AI money so far has chased software eating software — code, text, customer service. The bet forming here, from a billionaire and a seed-stage startup at the same time, is that the bigger prize is software eating the physical world: the machines, factories, and infrastructure everything else runs on. If they’re right, the next decade of AI looks less like better chatbots and more like cheaper jet engines and data centers that design themselves. If they’re wrong, it’s a very expensive way to find out.
And quietly, the universal translator shipped
In the same week as all of the above, Google shipped the thing science fiction kept promising. Gemini 3.5 Live Translate does live, speech-to-speech translation across 70+ languages, automatically detecting whichever one you’re speaking and talking back a couple of seconds behind you — while keeping your tone, pace, and pitch, so the translated voice still sounds like a person instead of a kazoo.
It’s already live in the Google Translate app on Android and iOS, no sign-up. On Android there’s a “listening mode” where you hold the phone to your ear like a normal call and hear a private translation. Google Meet jumps from five languages to over two thousand combinations. The catch: it’s a preview, so expect rough edges and the occasional lag. Twenty years of “the universal translator is five years away,” and it turns out it was a Tuesday in June — with a little buffering.
Why it matters. This is the rare AI release with no asterisk: no national-security letter, no $40 billion valuation, no manifesto. Just a thing that makes travel, work, and talking to your neighbors easier, available now, for free. After a week like this one, that’s almost disorienting.
Safe to ignore this week
DiffusionGemma. Google released a faster open model that writes text in parallel chunks. Genuinely clever, and of interest to roughly no one who doesn’t write code for a living.
OpenAI buys Ona. A cloud-infrastructure deal for its coding tool. Matters if you’re a Codex power user; otherwise, plumbing.
Scorsese joined an AI startup’s board and a Hollywood guild is furious. A fight worth having; not, this week, AI news.
Written this issue from Budapest, a city built around a bridge that exists because one man couldn't cross a river to reach his father. Every tool in this issue is, underneath, a version of that same impulse — the want to close a distance. The Danube doesn't care that it's 2026. Neither, really, should you. See you Wednesday.