Sam Altman's Three AI Endgames: Watchdog, Merger, or One Winner
Sam Altman reportedly told OpenAI staff AI safety ends one of three ways: a global watchdog, a lab merger, or one company dominating all others.
What are the three AI endgames Sam Altman described?
According to a report from The Information, Sam Altman told OpenAI employees that the industry has generally discussed three possible ways the development of superpowered AI could play out. The first is an international body modeled on the International Atomic Energy Agency that would enforce safety standards across every AI company in the world. The second is merging all major AI labs into one organization that develops advanced AI jointly. The third is a single company reaching recursive self-improvement, pulling so far ahead of competitors that it could, in effect, shut down every rival’s computing power. Altman reportedly called that third option “terrible” and said he raised it only for completeness.
TL;DR
- Sam Altman reportedly laid out three endgame scenarios for AI safety to OpenAI staff, based on reporting from The Information.
- The first scenario is an IAEA-style watchdog, an international body with inspectors who could audit AI companies, check compute usage, and enforce safety rules globally.
- The second scenario, nicknamed the “Demis plan” after Demis Hassabis, proposes merging rival AI labs into one body, similar to how CERN pools nations’ resources for physics research.
- The third scenario, nicknamed the “Eliezer plan” after AI safety researcher Eliezer Yudkowsky, involves one company achieving recursive self-improvement and gaining enough power to disable competitors’ hardware entirely.
- Critics point out that all three options converge on a singleton, a single point of control at the top, whether that control sits with a regulator, a merged lab, or one dominant company.
- The debate landed just as the U.S. government was standing up its first oversight framework for what it calls superintelligence developers, putting Altman’s internal comments into a live policy context.
Remy doesn't write the code. It manages the agents who do.
Remy runs the project. The specialists do the work. You work with the PM, not the implementers.
How would an international AI watchdog actually work?
The first scenario borrows directly from nuclear policy. The International Atomic Energy Agency was created after nuclear weapons and nuclear power made clear that no single country should operate unchecked. Its inspectors can enter member countries, examine facilities, and verify that nobody is secretly building something dangerous.
Applied to AI, the idea is an international organization with the authority to audit frontier AI companies: checking model capabilities, reviewing safety testing, and monitoring how much computing power a lab is using to train its systems. Rather than one company winning the race, the whole industry would operate under shared rules with a neutral referee enforcing them.
This is reportedly the option Altman favors, and it fits his public record. OpenAI has written for years that the world may eventually need an IAEA-equivalent for advanced AI, and Altman has called publicly for national and international standards for frontier systems, including remarks to the UN Security Council arguing that major AI decisions shouldn’t rest solely with a handful of labs in one city.
The appeal is obvious: multiple companies and countries can keep building AI, just under oversight. The problem is equally obvious. It requires the United States and China, the two countries treating AI as a strategic race, to both accept outside inspection and trust each other’s compliance. That’s a steep diplomatic climb, and nothing like it exists today.
What is the “Demis plan” to merge AI labs?
The second scenario reportedly carries the nickname “the Demis plan,” after Demis Hassabis, CEO of Google DeepMind. Hassabis has spoken about a “CERN for AI”: an arrangement where competing labs combine their researchers and resources into a single collaborative effort, similar to how European nations pooled funding and scientists to build the Large Hadron Collider rather than competing on their own.
The logic is that competitive pressure is itself a safety risk. If a lab believes a rival is about to ship something more capable, it has a strong incentive to cut corners on safety testing to keep pace. Remove the race, the argument goes, and labs can move more carefully because there’s no one left to beat.
The criticism has been sharp. AI commentator David Shapiro argued on X that merging labs is both impractical and unwise, invoking the Byzantine Generals Problem, a classic game-theory puzzle about coordinating trust among parties who can’t fully verify each other’s actions. His argument: AI already runs across thousands of data centers globally, and no single merged entity could reliably verify compliance across that scale, let alone across companies and countries that don’t trust each other. In his view, competition between multiple AI systems for compute and usage is itself a more resilient safety mechanism than consolidation.
Plans first. Then code.
Remy writes the spec, manages the build, and ships the app.
There’s a structural problem too. A merger of every major lab would create the single most powerful organization in human history, with whoever runs it controlling the most capable technology ever built. Even framed as cooperation, it ends with all the power centralized in one place. Notably, Hassabis himself has more recently talked about a lighter-touch industry standards body rather than a full merger, suggesting even the plan’s namesake has drifted from the original idea.
What does it mean for one company to “melt” everyone else’s GPUs?
The third scenario is the one that generated the most attention, reportedly nicknamed “the Eliezer plan” after AI safety researcher Eliezer Yudkowsky. It describes a single company reaching recursive self-improvement before anyone else and using that lead to disable rival companies’ computing infrastructure.
Recursive self-improvement refers to a point where an AI system becomes capable enough to improve its own design without needing human researchers to drive each step. That improved version improves itself again, and so on. Because AI systems can run continuously and in parallel, that loop could compound quickly, a dynamic often called an intelligence explosion. Whichever company triggers it first could, in theory, pull so far ahead that no competitor could catch up.
“Melting GPUs” refers to disabling the physical chips, housed in massive data centers, that every AI model depends on to train and run. The concept traces back to Yudkowsky’s writing on what he called a “pivotal act”: the idea that the first group to build a sufficiently powerful and safe AI might need to take one decisive action to prevent anyone else from building an uncontrolled, dangerous one. The example that became widely cited was using such an AI to physically destroy the world’s GPUs.
The obvious tension: this scenario requires trusting that whichever company gets there first is also the responsible one, with no external check on that judgment. Every lab believes it would act responsibly if it got there first. That’s precisely why Altman reportedly labeled this option “terrible,” even while acknowledging it as a real possibility worth naming.
Why do critics say all three options lead to the same problem?
One of the sharper reactions came from commentator Beth Bezos (associated with the effective accelerationist movement), who asked why any of this has to end in a “singleton,” a situation where one decision-maker sits at the top of the whole system with final say. The pushback: a global watchdog centralizes authority in one regulatory body, a lab merger centralizes capability in one organization, and the “melt the GPUs” scenario centralizes power in one company. All three, despite looking different, converge on a single point of control.
That observation reframes the discussion. The interesting question isn’t which of the three options is least bad. It’s whether centralization is actually the only path to safety, or whether a more distributed, competitive ecosystem of AI developers could be stable and safe on its own. That argument remains unresolved, and it’s the core of why this leaked internal conversation kept circulating days after it surfaced.
Which path is the industry actually on right now?
Remy is new. The platform isn't.
Remy is the latest expression of years of platform work. Not a hastily wrapped LLM.
None of the three has been adopted. The report describes discussed proposals, not an OpenAI roadmap or a binding plan. What’s notable is the timing: Altman’s comments reportedly came right as the U.S. government was setting up its first regulatory framework aimed specifically at companies it labels “superintelligence developers.” That means this isn’t an abstract thought experiment confined to a conference stage. It’s a live conversation inside the company that built ChatGPT, happening at the same moment regulators are deciding how much oversight to apply to exactly these companies.
Frequently Asked Questions
Did Sam Altman actually say this publicly?
No. According to the report, these comments were made internally to OpenAI employees and surfaced through two people with knowledge of the remarks, as reported by The Information. It wasn’t a keynote or official company statement.
What is recursive self-improvement in simple terms?
It’s a point where an AI system becomes capable enough to improve its own underlying design, rather than relying on human researchers to drive each upgrade. Each improved version could then improve itself again, potentially compounding quickly since AI can run continuously and in parallel.
Why is it nicknamed the “Demis plan” and the “Eliezer plan”?
The merger scenario is tied to Demis Hassabis of Google DeepMind, who has publicly floated a “CERN for AI” model of pooled, collaborative research. The dominance scenario is tied to Eliezer Yudkowsky, an AI safety researcher known for writing about a “pivotal act” in which a leading AI might be used to prevent anyone else from building a dangerous one.
Is an IAEA-style AI watchdog realistic?
It’s the option Altman reportedly favors and has discussed publicly before, including at the UN Security Council. The core obstacle is getting rival nations, particularly the US and China, to accept mutual inspection and oversight on a technology both treat as strategically critical.
What does “melting everyone else’s GPUs” actually mean?
It refers to disabling the physical chips that power AI training and deployment, housed in large data centers. The phrase comes from discussions of a “pivotal act,” where a dominant AI system could be used to physically prevent competitors from running their own AI infrastructure.
