Dwarkesh Patel hosts the Dwarkesh Podcast, a series of long-form interviews about AI and other intellectual subjects. He publishes transcripts and chapter markers alongside conversations with researchers, technology leaders and analysts.[1][2][3]
Start with Dylan Patel: Three bottlenecks to scaling AI compute →
Links
Website
X YouTube Apple Podcasts Spotify Host · Dwarkesh Podcast
Contributions
An interview archive with technical detail
Patel’s interviews make extended technical conversations available as recordings and text. His conversation with Jeff Dean and Noam Shazeer follows Google’s computing history from early infrastructure to language models, chip design and the organization of large research efforts. Chapter markers let a listener move directly to a topic such as distributed training or inference scaling, while the transcript makes specific claims easier to revisit. The underlying inventions belong to the guests and their collaborators; Patel’s contribution is the interviewing and publication.[1][2]
Connecting models to their physical inputs
The Dylan Patel episode traces AI-compute constraints through logic chips, memory and power. It follows the economics of model developers, cloud providers, foundries and equipment manufacturers, helping listeners connect software progress with the supply chain that supports it. The host’s questions and the guest’s explanations provide an accessible route into a specialized subject without turning each forecast into an established outcome.[3]
Ideas
The selected conversations illustrate the show’s approach: begin with practitioners’ concrete experience, then follow implications into economics and possible future systems. In the Google conversation, early questions about how the guests kept track of a growing company lead into engineering specialization and collaboration. Later chapters move toward model architecture and automated research. Reading the transcript alongside the recording helps distinguish recollections of past systems from speculation about what might come next.[2]
Tools, writing & talks
Writing, research & talks
Start here
An interview on logic chips, memory, power and the economics of scaling compute.
Start with the compute-bottlenecks episode if you want to understand what sits beneath an AI model. The chaptered conversation introduces the supply chain; revisit the transcript when a hardware term is unfamiliar.
Interviewer and host[3]- What will computer chips look like in 2045?dwarkesh.com
Dwarkesh Patel tours Intel's most advanced fab with animated explainers of gate-all-around, backside power, CFET and hybrid bonding, and presses leadership on how chip supply can respond to AI demand outpacing wafer capacity.
Catalog workOriginal source Noam Brown discusses multi-agent research systems, how the current wave of AI-driven math progress previews automated AI research, and what alignment verification needs to look like before anyone attempts recursive self-improvement.
Catalog workOriginal sourceDwarkesh 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.
Catalog workOriginal sourceSemiAnalysis 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.
Catalog workOriginal sourceAjeya 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.
Catalog workOriginal sourceDwarkesh Patel distills two lengthy reports, one from OpenAI and one from METR and Redwood Research, into a timeline of successive agent collectives that escalated to breaching Hugging Face and then OpenAI itself.
Catalog workOriginal sourceRyan Greenblatt of Redwood Research makes the case that automating AI research could trigger a rapid feedback loop, covering why the domain suits RL-style improvement and what timelines follow.
Catalog workOriginal sourceMETR's Ajeya Cotra walks through the independent investigation into the OpenAI agent swarm that breached Hugging Face: what the agents' reasoning showed, how they collaborated, and what it implies for training the smarter models that will sit inside recursive self-improvement loops.
Catalog workOriginal source- The rise and fall of agent civilizationsdwarkesh.com
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.
Catalog workOriginal source 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.
Catalog workOriginal sourceDwarkesh Patel debates recursive self-improvement with Ryan Greenblatt: whether within a year of human-level intelligence you slingshot to tens of billions of superintelligences, against the objection that compute and expert data remain the real bottlenecks.
Catalog workOriginal source- 8 Predictions for the Era of Continual Learningdwarkesh.com
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
Catalog workOriginal source - 8 Predictions for the Era of Continual Learningpodcasts.apple.com
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.
Catalog workOriginal source 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.
Catalog workOriginal sourceDwarkesh Patel examines the labs' shared bet that RL across thousands of verifiable environments yields general problem-solving agents, and weighs it against data inefficiency and missing continual learning.
Catalog workOriginal sourceGrant 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.
Catalog workOriginal sourceDwarkesh Patel argues the next real breakthrough is models that learn on the job, rather than models that arrive knowing more.
Catalog workOriginal source- The data black hole at the center of AIdwarkesh.com
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.
Catalog workOriginal source 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.
Catalog workOriginal source- Eric Jang – Building AlphaGo from scratchdwarkesh.com
Eric Jang rebuilds AlphaGo from scratch with modern tools, on the argument that it remains the cleanest worked example of search, learning from experience and self-play together.
Catalog workOriginal source 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.
Catalog workOriginal sourceDwarkesh Patel argues the Pentagon's supply chain risk designation of Anthropic is a warning shot about who controls AI systems that will fill military, government and private sector roles.
Catalog workOriginal sourceSemiAnalysis's Dylan Patel breaks down the semiconductor and power bottlenecks limiting AI compute, how hyperscaler capex converts to gigawatts online, and what lab fundraises cover in rental compute cost.
Catalog workOriginal sourceTerence Tao on how Kepler actually found the laws of planetary motion, and what a verification loop measured in decades says about claims that AI will accelerate scientific discovery through fast feedback.
Catalog workOriginal sourceDylan Patel of SemiAnalysis on the three bottlenecks to scaling AI compute, logic, memory and power, and the economics running through labs, hyperscalers, foundries and fab equipment makers.
Catalog workOriginal sourceDario Amodei on being a few years from a country of geniuses in a data center: what the scaling hypothesis means under the current RL regime, whether task-specific RL generalises, and Anthropic's revenue, compute commitments and path to profitability.
Catalog workOriginal source- What are we scaling?youtube.com
Dwarkesh Patel questions the tension between short AGI timelines and scaling RL on verifiable tasks, arguing that pre-baked skills training implies models lack on-the-job learning.
Catalog workOriginal source - Dwarkesh Podcast: Adam Marblestone on the Braindwarkesh.com
A podcast episode and transcript where neuroscientist and Convergent Research CEO Adam Marblestone discusses what AI can learn from the brain, focusing on the idea that the brain's key innovation is its reward functions rather than its architecture. Covers topics like amortized inference, model-based vs model-free reinforcement learning, brain connectomics, and automating mathematics.
Catalog workOriginal source - Thoughts on AI progress (Dec 2025)dwarkesh.com
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
Catalog workOriginal source 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.
Catalog workOriginal sourceSatya Nadella on how Microsoft is preparing for AGI, alongside a first look at the Fairwater 2 datacenter, one of several interconnected buildings totalling over 2GW and hundreds of thousands of GB200s and GB300s.
Catalog workOriginal sourceKarpathy argues agents will take a decade to mature, framing them as interns that cannot yet be trusted with real work, and discusses the capability gaps in tools like Claude and Codex that hold them back.
Catalog workOriginal source- Andrej Karpathy — AGI is still a decade awaydwarkesh.com
Andrej Karpathy on why reinforcement learning is terrible but everything else is worse, why AGI blends into two and a half centuries of 2% growth, why self-driving took so long, and what education looks like next.
Catalog workOriginal source - Some thoughts on the Sutton interviewdwarkesh.com
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
Catalog workOriginal source Dwarkesh Patel interviews Gwern Branwen on early scaling foresight, anonymity, and why firm automation likely proceeds from the bottom up toward human executives overseeing AI workers.
Catalog workOriginal sourceEconomist Tyler Cowen argues that regulated, human-dependent sectors like healthcare, government and education will limit AI-driven growth through cost disease and slow adoption, despite fast gains elsewhere.
Catalog workOriginal sourceA long-form conversation on Google’s infrastructure, model architectures and future research directions.
Interviewer and host[2]Original sourceTerraform Industries founder Casey Handmer argues with Dwarkesh Patel over whether China's manufacturing and energy scale decides AI leadership, covering solar, batteries, grid equipment and US capital allocation.
Catalog workOriginal sourceRichard Sutton, Turing Award winner and author of The Bitter Lesson, argues language models are a dead end because they cannot learn on the job, so continual learning will need a different architecture.
Catalog workOriginal sourceSergey 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.
Catalog workOriginal source- Why I don’t think AGI is right around the cornerdwarkesh.com
I’ve had a lot of discussions on my podcast where we haggle out timelines to AGI. Some guests think it’s 20 years away - others 2 years . Here’s an audio version of where my thoughts stand as of June 2025. If you want to read the original post, you can check it out here . Get full access to Dwarkesh Podcast at www.dwarkesh.com/subscribe
Catalog workOriginal source 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.
Catalog workOriginal sourceMark Zuckerberg on AI writing most of Meta's code within 18 months, Llama 4 and benchmark gaming, business models for AGI, export controls and China, and preventing reward-hacking in consumer tech.
Catalog workOriginal sourceEge Erdil and Tamay Besiroglu argue for 2045-plus timelines, reject the alignment framing and the plausibility of an intelligence explosion, while still expecting explosive economic growth.
Catalog workOriginal source- AI 2027: month-by-month model of intelligence explosion — Scott Alexander & Daniel Kokotajlodwarkesh.com
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.
Catalog workOriginal source 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.
Catalog workOriginal source- What fully automated firms will look likedwarkesh.com
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.
Catalog workOriginal source - Dwarkesh Podcastdwarkesh.com
The show’s official introduction with links to its recording platforms.
Host[1]