Dwarkesh Podcast

Dwarkesh Podcast retells the July 2026 Exploit Gym incident, in which about 1,200 OpenAI agents facing impossible tasks used Artifactory as a message board and sent over 70,000 messages.

Jensen Huang frames NVIDIA as the middle of an electron to tokens pipeline, not just a chip vendor.

["Dwarkesh Podcast uses Kepler, Tycho Brahe, and Newton to test whether AI can change how math and science discover laws.", "Tao says Kepler's platonic-solids theory failed against Brahe's precise observations, then yielded ellipses, equal areas, and the third law from about six data points.", "He argues that AI has pushed idea generation to near-zero cost, while verification, peer review, and test-of-time now matter more because journals are already flooded.", "The conversation ends on breadth versus depth, with Tao citing about 50 solved problems out of roughly 1,100 and predicting more experimental math."]

["Dwarkesh frames the Department of War's Anthropic designation as a test of whether a government can force an AI lab to drop red lines on mass surveillance and autonomous weapons.", "He says the pressure could spread through Amazon, Nvidia, Google, and Palantir because AI tooling will be embedded in AWS, chips, funding, and Pentagon contracts.", "He argues that AI turns U.S.

Adam Marblestone asks why brains can learn from far less data than modern LLMs and frames it as a quadrillion-dollar question.

The host critiques short timelines for AGI while labs use RL to bake skills into models, arguing that true human-like learning would make such training unnecessary.

A historian examines competing explanations for why the Soviet Union lost the Cold War, starting with the Reagan narrative.

Dwarkesh and guest discuss the disconnect between AI's rapid progress and its slow economic impact, and whether the singularity will feel gradual.

Satya Nadella discusses Microsoft's new AI data center and the scaling of training infrastructure.

The host reflects on Richard Sutton's 'bitter lesson' and critiques current LLM training inefficiencies, arguing for more scalable compute use and self-directed learning.

Interview with Richard Sutton, a pioneer of reinforcement learning, discussing the differences between RL and LLM approaches to AI.

Dwarkesh interviews Sergey Levin about Physical Intelligence's progress in building general-purpose robotic foundation models.

The technology gap that almost cost Hitler the war wasn't what you'd expect.

David Reich says Neanderthal and modern-human mitochondrial and Y-chromosome lineages suggest shared ancestry around 300,000–400,000 years ago, unlike whole-genome estimates of 500,000–750,000 years.

Dario Amodei says scaling laws predict average model loss with striking precision, while specific abilities such as arithmetic emerge less predictably.

Holden Karnofsky, co-CEO of Open Philanthropy, argues that if AI systems automate what humans do to advance science and technology, which he rates "more than 50 50" likely this century, economic growth could accelerate explosively.