Altman described three AI safety endgames to OpenAI staff, per The Information: an IAEA for AI, a CERN-style merger, and one firm melting rivals’ GPUs. Where each idea comes from.