Gonçalo Esteves / building capable systems Field Notes →

The Vault / People

Person

Paul Christiano

Created 2026-11-14, promoted off the Moonshots 311 — Jensen's Departure, NVIDIA's succession plan, AI doomer response, and Demis talks RSI (Live) decomposition.

status · researched added 14 Nov 2026 updated 14 Nov 2026 from Moonshots #288 · Moonshots #291 · Moonshots #292 · Moonshots #311 personai-safety see it in the graph →

Created 2026-11-14, promoted off the Moonshots #311 decomposition.

Who he is (one line): AI alignment researcher, co-author of RLHF (reinforcement learning from human feedback); led AI safety at the US AI Safety Institute (2024); joined the OpenAI Foundation board and its Safety & Security Committee (announced 9 Sep 2026).

Why it is in this archive

Gate: four held episodes, three as a segment subject. The regex index misses him because the captions spell him "Cristiano". Moonshots #288 (his move to the OpenAI nonprofit board; Alex: "Is it a revolving door or a trap door?"), Moonshots #291 (his personal statement read at length; Alex: "old Paul… should meet new Paul"), the 09-19 Tenev episode (his "18 months" to automated AI research, cited), Moonshots #311 (his Hinton Lecture).

The facts (checked 2026-11-14)

  • Hinton Lecture, Toronto, 9 Nov 2026: "I think there is a high risk of an AI takeover within the next three years", meaningful risk "even in the next 6 to 12 months"; takeover = AI systems pursuing goals and calling the shots with humans unable to stop them; "Any number we give is an expression of someone's subjective opinion"; on the fastest possible trajectory, "something like a coin flip" (BetaKit; clip on EP311 0:40:20–0:42:10).
  • Earlier number for context: 10–20% chance of a takeover with most humans dead (Bankless, 2023). Not a restatement.

What the panel said

episode who what ts
Moonshots #291 Alexander Wissner-Gross "I think old Paul who was advocating for prosaic alignment should meet new Paul"; "alignment equals capabilities" 0:56:29
Moonshots #311 Emad Mostaque "I don't really buy it at all like I have a high p doom but… what is the evidence"; "we don't have enough robots" 0:42:30–0:43:12
Moonshots #311 Alexander Wissner-Gross "I agree more with the old Paul"; "I would encourage Paul not to spread fear, not… to be a fear-monger" 0:45:02, 0:46:38
Moonshots #311 Salim Ismail "where do you calculate the downside risk of… inaction?"; "You're on the board of the thing accelerating the whole damn thing" 0:47:11, 0:48:16

Consistent across episodes: the panel's answer to him has not changed since September (old Paul vs new Paul; alignment = capabilities; P(fab)). His own estimate has sharpened (from OpenAI's quoted "18 months" to a three-year takeover window). No mate argues his side on any of the three tapes.

How this page stays current

Grep new transcripts for cristiano|christiano. Diff: any new number from him; any mate who moves toward his view.

Related

OpenAI · Recursive self-improvement (RSI) · Demis Hassabis · Alexander Wissner-Gross · Moonshots #311

Reading rule. Anything marked unverified or on-air is a claim made on the podcast and has not been checked against a primary source yet. It stays marked until it has. Sourced numbers link to where they came from.