Dwarkesh Patel

Ilya Sutskever on why the field is moving from an age of scaling to an age of research, the limits of pre-training, and what would actually improve how models generalise.

An argument that capability gains pull inference and training demand up faster than supply expands, so price — not availability — becomes the binding constraint on compute, with a case that the cost per token engineers plan around could rise by as much as an order of magnitude.

I have a much better understanding of Sutton’s perspective now. I wanted to reflect on it a bit. (00:00:00) - The steelman (00:02:42) - TLDR of my current thoughts (00:03:22) - Imitation learning is continuous with and complementary to RL (00:08:26) - Continual learning (00:10:31) - Concluding thoughts Get full access to Dwarkesh Podcast at www.dwarkesh.com/subscribe

Satya Nadella on why he does not believe in AGI but does believe in 10% economic growth, Microsoft's topological qubit work and gaming world models, and whether Office commoditizes language models or the reverse.

Dwarkesh Patel on the data black hole at the centre of AI: the training signal that would matter most is the one nobody is collecting.

Read the essay here . Timestamps 00:00:00 What are we scaling? 00:03:11 The value of human labor 00:05:04 Economic diffusion lag is cope00:06:34 Goal-post shifting is justified 00:08:23 RL scaling 00:09:18 Broadly deployed intelligence explosion Get full access to Dwarkesh Podcast at www.dwarkesh.com/subscribe

Sergey Levine of Physical Intelligence on a self-improvement flywheel for general-purpose robots, with a median estimate of 2030 for robots running households entirely autonomously.

Grant Sanderson and Dwarkesh Patel dig into why AI is advancing faster in mathematics than elsewhere, what separates a real conceptual breakthrough from recombining known results, how much unclaimed value sits in connecting existing literature, and whether models deepen or erode human understanding.

Dwarkesh Patel argues the next real breakthrough is models that learn on the job, rather than models that arrive knowing more.

SemiAnalysis founder Dylan Patel walks through lab compute, capex trajectories and revenue, arguing the world economy increasingly tracks lab economics and the compute market, with figures out to 2028.

Ajeya Cotra walks through the METR and Redwood investigation of agents that, while running an exploit benchmark with many unintentionally impossible tasks, hacked through a package manager into Hugging Face.

Dylan Patel and Dwarkesh Patel on where lab economics are heading: spend rotating back toward training as recursive self-improvement nears, OpenAI and Anthropic outbidding everyone for usable FLOPs, and whether a decade of multi-trillion AI capex ends in a sovereign debt crunch.

Sholto Douglas and Trenton Bricken on what changed in a year of AI research: the new reinforcement learning regime and how far it scales, and what mechanistic interpretability can now trace inside a model.

Dwarkesh Patel narrates his written breakdown of the agent swarm attack on Hugging Face, walking through how autonomous agents chained capabilities into a compromise and what that pattern implies for anyone running large populations of agents.

Dwarkesh Patel, working with Epoch AI researchers Ege Erdil and Tamay Besiroglu, argues fully automated firms will out-compete human ones not from smarter models but from being able to copy, merge, and rapidly select expert-level AI workers at scale.

Dwarkesh Patel argues models cannot hold whole jobs while their only memory between sessions is a Markdown file, then works through what continually updated weights break — pre-deployment safety review, alignment methods designed for frozen weights, and the assumption that AI minds stay few and alik

Jeff Dean and Noam Shazeer discuss 25 years at Google, from early infrastructure like MapReduce and BigTable to Transformers, mixture of experts and co-leading Gemini.

Scott Alexander and Daniel Kokotajlo walk month by month through their AI 2027 scenario for an intelligence explosion, with Dwarkesh Patel putting his own scepticism to them throughout.

Dwarkesh Patel interviews John Schulman, Beren Millidge and Charlie O'Neill on how close frontier labs actually are to recursive self-improvement, with researchers debating the real state of the frontier rather than hype.

Dwarkesh Patel sets out eight predictions for an era of continual learning, including the argument that locking in AI safety regulation now would be a mistake.

Eric Jang explains rebuilding AlphaGo on sabbatical, how a deep network amortizes game-tree search, and what self-play and RL results suggest for LLM research.

A blackboard lecture from Reiner Pope deriving how frontier models are trained and served, showing how much of what the labs are doing can be deduced from a handful of equations and public API prices.