Paul Christiano
Alignment researcher, and for years the most prominent careful optimist in debates with Eliezer Yudkowsky: he accepts most of the case for AI risk but puts the odds of survival much higher. He co-wrote the 2017 paper that introduced reinforcement learning from human feedback (RLHF), the method behind ChatGPT. He led alignment work at OpenAI, founded the Alignment Research Center, became head of AI safety at the US AI Safety Institute in 2024, and joined the board of the OpenAI Foundation in 2026.
2021: The IMO bet with Yudkowsky
In a 2021 exchange on LessWrong, the two tried to find predictions they disagreed on. One was whether AI would win gold at the International Mathematical Olympiad by 2025: Christiano gave it about 8%, Yudkowsky at least 16%. AI systems from Google DeepMind and OpenAI reached gold-medal level in July 2025, so this one went Yudkowsky’s way. The bet mattered because it tested their real disagreement: whether AI progress comes smoothly and predictably, as Christiano expected, or in sudden jumps.
2022: Where he agrees and disagrees
Where I agree and disagree with Eliezer (2022-06-19) is his reply to AGI Ruin - A List of Lethalities. He agrees with much of it, but expects a slower, more gradual AI takeoff and more chances to learn from weaker systems.
2023: Timelines and risk
- 2023-04-27, My views on “doom”: about a 22% chance that AI takes over, and higher total odds of a bad outcome once other risks are included.
- 2023-10, Dwarkesh Podcast: about a 15% chance of AI capable of building a Dyson sphere by 2030 and 40% by 2040, that is, AI far past human level.
2026: OpenAI board
On 2026-09-09 OpenAI named him a director of its nonprofit foundation and a member of its Safety and Security Committee. In a statement on joining, he said the AI industry, OpenAI included, is not on track to reduce the risk of catastrophic loss of control to an acceptable level.
See AGI timelines.