The Best Dwarkesh Podcast Episodes: A Listener’s Guide (2026)

The Dwarkesh Podcast is the strangest success story in podcasting. It started in 2020 as The Lunar Society, a college student’s interview show with almost no audience. Six years later, The Economist calls Dwarkesh Patel “Silicon Valley’s favourite podcaster,” frontier lab CEOs give him hours of their time, and the show charts at the top of tech podcasts across dozens of countries.

The reason is preparation. Dwarkesh reads everything his guests have written, sometimes years of papers, and guests routinely say some version of “nobody has ever asked me that” on air. The result is a back catalog where the best episodes aren’t just interviews, they’re the clearest available records of what the people building AI actually believe.

That back catalog is also intimidating: episodes run two to four hours. So here’s the guide I wish I’d had, based on the show’s own most-popular rankings and too many hours of my own listening. (I put the show in my best AI podcasts guide as the one to pick if you only pick one; this is the episode-level version.)

Start with these three

1. Andrej Karpathy: “AGI is still a decade away” (October 2025)

The most-watched episode of the show for a reason. Karpathy, fresh off OpenAI and Tesla, spends two and a half hours explaining why he thinks the current systems are more impressive and less finished than either camp admits: agents that don’t learn on the job, reinforcement learning’s “sucking supervision through a straw,” and what a real decade of progress looks like. It reset the vibe of the entire AGI-timeline debate within a week of airing.

2. Ilya Sutskever: “We’re moving from the age of scaling to the age of research” (November 2025)

Ilya almost never talks. When the co-founder of OpenAI and now SSI broke his silence here, the episode was dissected line by line across the industry the same day. The core claim, that pure scaling has given way to a new research era, is the single most quotable thesis about where AI is in 2026.

3. Sarah Paine: the lecture trilogy (Japan, India, Russia)

The show’s biggest surprise: a Naval War College historian became a breakout star. Her lecture-plus-interview doubles on why Japan lost WWII, the war for India, and why Russia lost the Cold War have nothing to do with AI, and listeners rank them among the best things the show has ever aired. If you want to understand why people call this podcast a university, start with the Japan lecture.

The rest of the essential list

Satya Nadella (November 2025). Four and a half hours, including a walking tour of a Microsoft datacenter. The rare CEO interview where the CEO gets pushed on capital expenditure math and answers anyway.

Leopold Aschenbrenner (June 2024). The four-hour conversation that landed alongside his “Situational Awareness” essays and set the terms of the AGI-by-2027 debate. Whether you find it prophetic or overreaching, half the discourse since is downstream of it.

Jensen Huang. Nvidia’s founder on supply chains, TPU competition, and why he runs the company the way he does. Pairs perfectly with the Satya episode for a picture of the biggest infrastructure buildout in history.

Dario Amodei. Anthropic’s CEO on what the end of the scaling exponential means. Listen back to back with the Ilya episode; they disagree in instructive ways.

The solo essays. Dwarkesh’s own audio essays, like “Why I don’t think AGI is right around the corner” and his China notes, chart alongside the big interviews. He’s one of the few hosts whose monologues are worth the same attention as his guests.

How to actually listen to these

Dwarkesh episodes are dense in a specific way: they’re full of predictions with dates attached. Karpathy says a decade. Leopold says 2027. Dario, Ilya, Satya, and Jensen all stake positions that will look either brilliant or embarrassing by 2028.

Which makes this the single best show in podcasting for timestamped notes. When someone makes a claim, mark the moment: with Margin you press and hold, say “Karpathy: agents can’t learn on the job yet, that’s the bottleneck,” and the note pins to that exact second. Six months later you have receipts, and the receipts are where the fun is. My general system for this is in the marginalia method.

One warning from experience: don’t listen to these at 2x. I wrote about what speed does to comprehension; these are exactly the episodes that deserve the slack.

Selinay