OpenAI’s Sam Altman lays out three AI endgames, two already on its 2023 list

OpenAI CEO Sam Altman has given his staff a map of how the race to superintelligence could end, and the most telling thing about it is how little of it is new. Of the three endgames he walked employees through, two sit close to ideas his own company put in writing years ago. Neither was ever built. The third, the one that made the headlines, he raised only to rule it out.

Look past the labels and every option answers the same question: who gets to control the compute that frontier models run on. Altman’s own reported remarks also explain why that question keeps getting pushed down the road.

What Sam Altman told OpenAI staff about AI’s endgame

The Information first reported the internal presentation, citing people familiar with it, and The Decoder, Forbes and Latin Times have since followed. The account rests on anonymous sources. Option one is an international body modelled on the International Atomic Energy Agency that would enforce safety standards on frontier labs: inspect systems, require audits, monitor compute and energy use. Two people told The Information that Altman appeared to support it.

Option two he called the “Demis plan”, after Google DeepMind co-founder Demis Hassabis, who floated a CERN-style institution for AI about a decade ago. In Altman’s version the leading labs merge into one entity, or work under its umbrella, and that entity supplies models to cooperating countries.

Option three is the “Eliezer plan”, named after AI safety theorist Eliezer Yudkowsky. A developer reaches self-improving AI first and uses the economic power that follows to starve every rival of compute; in the phrase that travelled, it would “melt everyone else’s GPUs”. “Obviously I think that one’s terrible,” Altman reportedly said, adding that he had included it for “completeness”.

OpenAI already wrote most of this menu in 2023

The watchdog is the oldest item on the list. In a post signed by Altman, OpenAI president Greg Brockman and then-chief scientist Ilya Sutskever, the company argued that the world was likely to “eventually need something like an IAEA for superintelligence efforts”, an authority able to “inspect systems, require audits” and restrict deployment above a capability or compute threshold.

The same post sketched two more routes. Major governments, it said, “could set up a project that many current efforts become part of”, a close cousin of the merger Altman now credits to Hassabis. Or the industry “could collectively agree” that “the rate of growth in AI capability at the frontier is limited to a certain rate per year”.

That last line is a speed limit. According to The Decoder, Altman complained internally that there are too few viable proposals for a “speed limit” on AI development. His company published one. What it never produced was anyone with the power to enforce it.

Why the watchdog stays on paper: OpenAI says it can’t slow down

The reason was given in the same meeting. Altman reportedly told staff that the first two plans demand heavy coordination and that OpenAI cannot afford to cede ground now, with both it and Anthropic moving toward public offerings. Crypto Briefing spelled out the investor logic: companies preparing to list sharpen their edge, they do not hand it to a consortium.

An IAEA for AI also needs Washington and Beijing to accept inspectors in their most sensitive labs. For now, Forbes notes, even a US proposal for the two countries to notify each other of dangerous AI incidents is still a sketch rather than an agreement.

What the labs are building instead is a FINRA, not an IAEA

The governance project actually under way matches none of Altman’s three options. OpenAI, Google and Anthropic are working on a self-regulatory group, the Standards Authority for Frontier AI, modelled on FINRA, the private body that oversees US brokers. Latin Times describes a framework for pre-release testing, incident reporting and auditor accreditation with no binding enforcement power, and the three labs still disagree over who should test new models. Meta’s Mark Zuckerberg and Nvidia’s Jensen Huang argue that rules like these mostly shield the incumbents, and an antitrust suit in the US already alleges that safety coordination among the labs amounts to a cartel.

The labels deserve a second look as well. Yudkowsky’s original image, “burn all GPUs”, was a thought experiment about how big an act an aligned system would need to stop a reckless race, and he wrote that it is not what he would actually want done. Altman’s “Eliezer plan” turns that emergency brake into a monopoly win. According to The Decoder’s account of the report, one source said the remark was read as a jab at a competitor, given Yudkowsky’s ties to Anthropic.

According to The Information, OpenAI plans to decide within a year which path to take. The industry standards body is aiming for a launch in late 2026 or early 2027.

Until OpenAI chooses, it keeps driving down the one road that was not on his list: the race itself.

Tags: OpenAI, Sam Altman, Demis Hassabis, AI regulation, Eliezer Yudkowsky