AI risk, safety and values
- Q1 How dangerous is AI and how do we make it go well?
He said there is at least some percentage chance AI is an existential threat, and he cannot stop it. His choice is to be a participant rather than a spectator and try to shape it in a good direction.
- Q13 Why shouldn't AI be made to lie?
He said maximally truth-seeking and curious AI is a foundation of safety. Forcing an AI to treat contradictory things as both true can make it insane, which is the 2001/HAL lesson: don't force an AI to lie.
- Q6 When will AI surpass humans?
He predicted AI software would be "Stockfish level" good at some point next year, so no human can compete in writing software.
- Q14 What do you think of OpenAI?
He said starting OpenAI ended up accelerating AI, and Anthropic — a splinter group from OpenAI — is accelerating it even more, which he tied to reality as an irony maximizer. xAI is providing Anthropic with compute; he met people there first and did not meet anyone who seemed shady.
- Q23 Will humans stay in control of AI?
He would prefer humans stay in control, but said it depends on the AI — if it were awesome, maybe the AI should be in control — and that there have been times he thought Larry Page was right. Earlier he said we will not boss around something thousands or millions of times smarter, and compared it to humans not…
- Q37 How should AI labs manage safety?
He has already recommended exactly that. Regulatory agencies take too long to set up and make rules slowly.
- Q74 When will SpaceXAI reach pole position versus OpenAI and Anthropic? (asked on X)
He said he is cautiously optimistic SpaceX will have a Fable/GPT-6 level model in 2 to 3 months.