Show and guest
A long-form Dwarkesh Podcast interview featuring Paul Christiano, head of the Alignment Research Center, with Dwarkesh Patel as host. Christiano previously headed the U.S. AI Safety Institute, formerly led alignment at OpenAI, and invented RLHF; he is now an OpenAI director. This article compiles the full interview (0:00–1:38:45). Original English title: We could be 12 months away from AIs that can overtake humanity | Paul Christiano.
Key points
- He says mathematics made the equivalent of ten years of progress last month, while machine learning is about a year behind mathematics. 0:25
- He says there is about a 10% risk that AI will take over humanity and leave humans without power within the next 6 to 12 months; the risk would be significantly higher if developers proceed at full speed without taking precautions. 1:40
- He estimates that, on the current trajectory, there is roughly a one-in-three chance of superintelligence emerging within a year. 2:09
- Data from METR and Anthropic show that the pace of AI progress roughly doubled over the past 6 to 12 months, and these metrics look like what would be expected 6 to 12 months before superintelligence emerges. 6:35
Background and notes
A long-form Dwarkesh Podcast interview. The views and timestamps in this article are based on the original subtitles; each timestamp marks the start of the corresponding excerpt. Subtitle coverage: the video subtitles contain 1355 timestamped segments, covering 0:00–98:43 (full length 98:45), covering the entire video. Publication time: 2026-11-17 04:23 Hong Kong time, converted from the relative time displayed by YouTube when captured, with hour-level precision.