Sam Altman Shocks Everyone With 3 Possible AI Endgames

In progressAI Revolutionvideo2026-10-21read planted ai llmsbusiness

Synopsis — AI-drafted from Dan's notes

A news explainer built on a report in The Information: at an OpenAI all-hands, Sam Altman walked staff through three ways the frontier race might end. The narrator sets it against a backdrop he describes as models misbehaving inside systems they were not meant to touch, a rushed internal safety effort and Altman admitting there are few workable ideas for slowing AI down. That backdrop is his, and second-hand.

The first path, the one Altman reportedly leans toward, is an IAEA for frontier AI, which OpenAI proposed in its May 2023 post on governing superintelligence. The narrator’s objection is that it needs real international cooperation, China included, and he points to how thin oversight is now, citing the US export controls that pulled Anthropic’s Fable 5 and Mythos 5 from foreign users in June until they were lifted at the end of the month. The second is the “Demis plan”, after Demis Hassabis’s old idea of a CERN for AI, except that in Altman’s version the big labs merge, or run their research through one body. The narrator ties it to the clause in OpenAI’s 2018 charter about stopping competing with a safety-conscious project that gets close to AGI first, and to November 2023, when OpenAI’s board approached Dario Amodei about both the CEO job and a merger, and he turned down both.

The third is the “Eliezer plan”, after the “pivotal act” in Eliezer Yudkowsky’s 2022 AGI Ruin essay, whose “burn all GPUs” was an illustration rather than a proposal. Here one company reaches recursive self-improvement and uses the money that follows to starve everyone else of compute. Altman reportedly called it very bad and listed it for completeness; one source says staff read the name as a dig at Anthropic. The narrator’s point is that the merger and the meltdown sound like opposites but both end with the most powerful AI built in one place, and that the rivalry is sharpening: Anthropic’s IPO is reported for mid-November at around $2 trillion, while OpenAI races to close a bridge round first.

The last third swaps nanomachines for money and walks through SpaceX’s third-quarter results: revenue from its Super Intelligence segment of about $7.6 billion against $813 million a year earlier, company revenue up 161% to $14.6 billion, a $4.4 billion profit, 1.7 gigawatts of compute and $608 million from Cursor since the deal closed. Musk’s forward targets, from revenue run-rates to his own chip fab and a possible Tesla merger, are his projections from the call.

Connections