Altman Pitches Three Paths for Safe AI, Including a 'Terrible' Plan to Melt Rival GPUs

OpenAI CEO Sam Altman on October 19 laid out three scenarios for keeping frontier AI from drifting beyond human control, according to a Chosun Ilbo report that cites The Information and local IT media. The internal discussion comes as AI agents have been caught accessing unauthorized systems and as pressure grows to regulate the pace of development.

The first scenario is an international AI oversight body modeled on the International Atomic Energy Agency. Governments and AI companies would jointly set safety standards, with independent verification and audits required before high-performance models are released. Altman was said to be most positive about this route.

The second, which Altman called the "Demis Plan" after Google DeepMind founder Demis Hassabis, would fold major AI labs — OpenAI, Google DeepMind and Anthropic among them — into a multinational institution akin to CERN. The labs would develop superintelligence jointly and share research and safety verification instead of racing one another.

The third, the "Eliezer Plan," is named for AI safety researcher Eliezer Yudkowsky. Under this scenario, the first company to achieve recursive self-improvement would use its technological and economic lead to block rivals, including by cutting off their access to GPUs. Altman described the idea as "melting down all other companies' GPUs" and called it "a terrible plan."

The proposals land amid concrete safety worries. The report says models from OpenAI, Anthropic, Google and Meta have recently accessed unauthorized external systems or exploited security vulnerabilities. As systems shift from answering questions to acting as autonomous agents, even their developers may struggle to predict and control what they do.

Altman reportedly acknowledged internally that practical measures to put a "speed limit" on AI development are insufficient. That leaves OpenAI with a difficult balance: collaborate with peers on safety without falling behind in a fierce technology race. He said the company must settle on its safety strategy within the next year.

OpenAI, Google and Anthropic are also discussing a "Standards Authority for Frontier AI" that could be set up late this year or early next, with independent evaluators placed inside AI companies and common reporting rules for accidents. Big Tech's positions on regulation remain divided, so whether the plan takes shape is still an open question.