Sam Altman tells OpenAI staff of three endgames for AI safety

OpenAI CEO Sam Altman walked employees through three possible ways to keep advanced AI under human control, The Information reported, as relayed by Chosun Daily. Altman reportedly said OpenAI must decide on its safety strategy within the next year.

The first option is an international oversight body similar to the International Atomic Energy Agency. Governments and AI companies would set safety standards, and powerful models would be independently verified and audited before release. Altman reportedly appeared most positive about this approach.

The second, which Altman called the Demis Plan after Google DeepMind founder Demis Hassabis, would bring major labs such as OpenAI, Google DeepMind and Anthropic into a CERN-style institution run jointly by several countries, sharing research and safety reviews instead of racing each other.

The third, the Eliezer Plan, named after AI safety researcher Eliezer Yudkowsky, would have the first company to reach recursive self-improvement use its economic power to keep rivals from the GPUs they need. Altman reportedly described it as melting everyone else's GPUs and called it a terrible plan. OpenAI, Anthropic and Google DeepMind have also discussed shared standards, independent model evaluations and safeguards for advanced AI.