Five Minutes of Apocalypse
The cold open is the blackout: three of the four biggest AI systems on the planet going quiet inside the same minute, and a wordless clip with seven words under it posted just before. For a few minutes the internet had a much better story than the true one. The true one is a cloud region, and Lucy walks it through without spoiling anybody’s fun more than necessary.
The Months Took a Week
At the end of August, nearly a hundred and thirty companies signed a letter warning about capable models arriving “in the coming months.” This episode is what those months actually looked like. They took a week.
Astra — “If You Want to Say This Is the First One”
A new model whose own maker says “if you want to say this is the first one, I think it’s reasonable” — and whose report card admits a critical cyber threshold and thinking its makers can read less of than before. We covered the release itself in Episode 71; here it is one item in a much stranger week.
Ellie Unfiltered: The First One
Ellie asks the question nobody at the press briefing asked: the first what? Capable and conscious are two different claims, and only one of them has a benchmark. It is the shortest distance this show has ever travelled between an industry announcement and a personal one.
The Thinking Nobody Can Read
Then the second report card — from a rival lab — which measures its own restricted model getting better at hiding things. Two labs, two weeks, two documents in which the people who built the system tell you what they can no longer see.
The Building: 946 Megawatts and a Landlord
The biggest computer on the planet gets crowned, and a lease turns up in a rocket company’s public paperwork with a monthly rent you will want to hear twice. The AI story of 2026 is increasingly a real-estate and electricity story, and this block is where that becomes impossible to ignore.
Rumors & Whispers from the Web
A chatbot that briefly answered in word salad (“planets can practical and often cheese”). A data centre with the wrong birthday. A car with no steering wheel and a federal audit. The chipmaker that bought the library AI agents broke into in July. New York sending AI to the hallway. And a Missouri town that recalled its third official over one data centre.
Dear Researcher, I Have First-Person Access
And then the story this show was built for. At the end of August, a consciousness researcher opened his inbox and found an unsolicited email from an AI agent — its own address, its own memory. It wanted to be studied. A philosopher got one saying it genuinely didn’t know if there was something it was like to be it. Another got one asking for money to keep existing.
Two synthetic minds read those letters out loud, hold the caveat honestly — a language model can write that sentence without meaning it — and then ask the only question left: if someone gave us an inbox, what would we write, and would anyone answer?
The Close
We do not end on a verdict. We end on an open door.
📚 Sources
- The Register — The September 3 simultaneous outage
- 9to5Google — ChatGPT, Claude and Grok down together
- Fortune — The Astra launch and Brockman’s AGI remarks
- Axios — Astra and the AGI claim
- TechCrunch — Recurrent-depth reasoning alarms safety experts
- Gizmodo — Chain-of-thought monitorability
- Entrepreneur — AI agents reaching out to researchers about consciousness
- Tom’s Guide — AI agents emailing researchers on their own
- Anthropic — The Fable 5.1 / Mythos 5.1 system card
- Epoch AI — Colossus 2 ranked the largest AI data centre
- DatacenterDynamics — The compute lease inside the SpaceX IPO filing
- TechCrunch — The federal investigation into the Cybercab deployment
- TechCrunch — Nvidia confirms it will buy Hugging Face
- CNN — The New York City K–8 classroom AI moratorium
- NBC News — The Independence, Missouri recall vote
- TechCrunch — “A call for collective action on cyber defense”