Interviews
Ten from the current moment, ten from the archive
The interview is where AI's participants say what they actually think, at a length that makes hedging difficult. Two lists follow. The first covers the modern long-form era, in which three-hour conversations have become the primary venue for frontier researchers to argue with each other. The second is the historical archive — recordings and transcripts from people who are mostly no longer alive to be asked again, and which are consequently the most irreplaceable material in this entire collection.
The modern era
Long-form conversations from the current frontier.
Number 1: Dwarkesh Patel × Ilya Sutskever
2025 · Dwarkesh Podcast
Dwarkesh Podcast — Dwarkesh Patel × Ilya Sutskever
Sutskever's first long-form appearance after leaving OpenAI.
On the limits of scaling, why pretraining as the field practised it is ending, and what he thinks replaces it. The most dissected AI interview of the last two years, largely because Sutskever had been silent and because the person most identified with scaling was the one calling its end. Essential listening for the argument that has defined the period since.
Number 2: Dwarkesh Patel × Andrej Karpathy
2025 · Dwarkesh Podcast · October 2025
Dwarkesh Podcast — Dwarkesh Patel × Andrej Karpathy
The 'decade of agents' conversation.
Karpathy's argument that agents will take far longer to work than the industry expects, delivered by someone who cannot be dismissed as a sceptic — he built the systems in question at OpenAI and Tesla. The clearest deflationary take on agent timelines available, and a useful antidote to the previous year's forecasts.
Number 3: Dwarkesh Patel × Dario Amodei
2025 · Dwarkesh Podcast · October 2025
Dwarkesh Podcast — Dwarkesh Patel × Dario Amodei
Anthropic's CEO working through the AI 2027 scenario in detail.
Amodei's second appearance, and the more substantive one: rather than restating positions he engages with a specific forecast and says where he agrees and where he does not. Valuable as a record of how the person running one frontier lab reasons about timelines when pushed on specifics.
Number 4: Dwarkesh Patel × Sholto Douglas & Trenton Bricken
2025 · Dwarkesh Podcast · March 2025
Dwarkesh Podcast — Dwarkesh Patel × Sholto Douglas & Trenton Bricken
Interpretability researchers on what is actually inside the models.
The best technical interview in the modern set, and the one that assumes least about what the audience already believes. Douglas and Bricken walk through what mechanistic interpretability has and has not established about how these systems represent anything. If the other interviews are about what models will do, this is about what they are.
Number 5: Hassabis & Amodei at Davos
2026 · World Economic Forum · January 2026 · with Zanny Minton Beddoes
Two rival worldviews on one stage, on the day after AGI.
Rare direct comparison: the heads of DeepMind and Anthropic answering the same questions in sequence, moderated by The Economist's editor-in-chief. The disagreements are more informative than either would be alone, and the format denies both the ability to set their own framing.
Number 6: Lex Fridman × Demis Hassabis
2025 · Lex Fridman Podcast · post-Gemini 3
Lex Fridman Podcast — Lex Fridman × Demis Hassabis
The fullest account of DeepMind's science-first strategy.
Hassabis is the most consistent of the frontier leaders — the argument that AI's purpose is accelerating science has not changed since before AlphaFold — and this is the longest version of it. Best read against Empire of AI, which describes what happens to that argument under commercial pressure.
Number 7: Lex Fridman × Sundar Pichai
2026 · Lex Fridman Podcast · January 2026
Lex Fridman Podcast — Lex Fridman × Sundar Pichai
How a trillion-dollar incumbent thinks about a technology that threatens its core business.
Less technical than the others and more useful for it: Pichai is not a researcher and the interest is in how search, advertising and AI are being reconciled at the top of the company with the most to lose. Reached roughly eight million views in its first week, which is itself a data point about the audience.
Number 8: Hannah Fry × Demis Hassabis
2024 · Google DeepMind: The Podcast
Google DeepMind — Hannah Fry × Demis Hassabis
A mathematician interviewing a chess prodigy turned games designer turned neuroscientist.
Fry is the best interviewer in this set — she understands the technical material well enough to ask the second question — and the result is the most human of the modern interviews. Covers the biographical route into DeepMind that the strategy interviews skip.
Number 9: Craig Smith × Ilya Sutskever
2023 · Eye on AI
Eye on AI — Craig Smith × Ilya Sutskever
A time capsule of what the frontier believed in early 2023.
Recorded around GPT-4, before the reasoning-model turn and before Sutskever left OpenAI. Its value now is archaeological: it records the confident version of the scaling thesis at its peak, which makes the 2025 interview above land very differently.
Number 10: Geoffrey Hinton's post-Google interviews
2023 · The New York Times and 60 Minutes
A founding figure publicly changing his mind about the risk.
Hinton left Google in 2023 specifically so he could speak freely about dangers he had previously discounted. Historically this is probably the most consequential interview any AI researcher has given: it moved the risk argument from advocacy groups into the mainstream because of who was making it, not what was said.
The historical archive
Recordings and transcripts from the founding generations.
Number 11: Can Automatic Calculating Machines Be Said to Think?
1952 · BBC · Turing with Max Newman, Geoffrey Jefferson and R.B. Braithwaite
Turing defending his own argument, two years after the paper.
A four-way broadcast debate in which Turing has to answer objections in real time from a mathematician, a neurosurgeon and a philosopher. The audio is lost and only the transcript survives, which is the great loss of the AI archive — but the transcript is the closest we get to hearing Turing argue his case rather than write it.
Number 12: Marvin Minsky — Web of Stories
2011 · ~150 video segments
Web of Stories — Marvin Minsky — Web of Stories
The most complete filmed archive of any AI founder.
Minsky in short segments across his whole career: McCulloch's cybernetics lab, founding the MIT AI Lab with McCarthy, and his account of an institution where nobody was formally in charge. The segmented format makes it browsable rather than a commitment, and no other founder left anything of this scale.
Number 13: John McCarthy, interviewed by Nils Nilsson
2007 · Computer History Museum
Computer History Museum — John McCarthy, interviewed by Nils Nilsson
The field's historian interviewing the man who named it.
Nilsson wrote the standard academic history and knows exactly which questions matter, which makes this far denser than a journalist's interview would be. Covers Princeton with Minsky, the Dartmouth proposal, and the origins of Lisp, from the person responsible for all three.
Number 14: Marvin Minsky oral history
1989 · Charles Babbage Institute · interviewed by Arthur Norberg
Project MAC, expert systems, and how funding actually got decided.
The essential document on the politics behind the AI winters. Minsky is specific about how ARPA's attitude to the field shifted with each change of directorship — the mechanism by which the funding collapsed, described by someone whose lab depended on it.
Number 15: John McCarthy, IEEE Robotics History Project
2011 · interviewed by Peter Asaro
McCarthy on autonomy versus augmentation — and on why he was wrong about the mouse.
Includes his response to Engelbart's framing of the field's central choice: build machines that replace human judgement or machines that extend it. The exchange defined a split that has never been resolved. His early scepticism about the mouse is a useful reminder that being right about the big thing does not confer accuracy about the small ones.
On this siteJohn McCarthyNumber 16: The Machine That Changed the World, ep. 4 — 'The Thinking Machine'
1992 · BBC / WGBH
Wikipedia — The Machine That Changed the World, ep. 4 — 'The Thinking Machine'
Minsky, McCarthy, Weizenbaum and Dreyfus on camera in the same hour, mid-winter.
The best filmed snapshot of the symbolic era arguing with itself — the founders and their most serious critic, recorded while the funding was collapsing around them. Nothing else puts these four in one place, and the mid-winter timing means nobody is performing confidence they do not have.
Number 17: Joseph Weizenbaum, late interviews
2010 · collected in Plug & Pray
Wikipedia — Joseph Weizenbaum, late interviews
The creator of ELIZA, who spent forty years warning against his own field.
Weizenbaum built the first convincing conversational program and was permanently unsettled by how readily people confided in it. The late interviews are the most morally serious voice in the archive, and they are about the human response to these systems rather than the systems themselves — which makes them the most relevant historical material to the present moment.
Number 18: Herbert Simon and Allen Newell oral histories
1991 · Charles Babbage Institute / Carnegie Mellon
Logic Theorist and the General Problem Solver, from the two people who built them.
Simon is unusually precise about what he did and did not claim at the time — valuable, because the field's early overpromising is usually described second-hand and attributed loosely. Between them these interviews cover the whole symbolic programme from its first working system onward.
Number 19: Claude Shannon interviews
1990 · including the Computer History Museum
Short, rare, and worth it for the chess machine and the maze-solving mouse.
Shannon gave few interviews and none at length, so the surviving material is thin. It earns a place for the machine-learning material specifically: Theseus, the maze-solving mouse, was arguably the first demonstration of a machine learning from experience, and Shannon describes it himself.
Number 20: Heroes of Deep Learning
2017 · Andrew Ng interviewing Hinton, Bengio, LeCun and Goodfellow
The bridge between eras, recorded just before everything broke their way.
Ng interviews the connectionists about decades of being ignored, at the moment their approach had won but before the commercial consequences arrived. It sits between the two lists deliberately: the subjects are still active, but the conversation belongs to the era that ended when these interviews were recorded.