Topics / AI Reasoning
AI Reasoning Podcasts
5 shows and 2,141 episodes sit under AI Reasoning.
Curated podcast guide to reasoning models — Noam Brown on test-time compute, Nathan Lambert on RLVR, the Cognitive Revolution on mathematical superintelligence
Photo: Brecht Corbeel on Unsplash · Unsplash License
Swipe · AI Reasoning only
Swipe through AI Reasoning
Save time discovering your next favourite podcast by listening to just the best minute of an episode.
The ai reasoning slate · every cover a link
All 5 ai reasoning shows →
Next to this on the shelf
Explore by subtopic
12 subtopics
A field guide
AI Reasoning podcasts, explained
Reasoning gets interesting on a podcast when the conversation moves past benchmark scores and asks what produced the answer. Sometimes that means test-time compute. Sometimes it means causal data, evaluation, backpropagation, or the awkward gap between a model’s apparent competence and its actual reliability.
The best episodes make room for uncertainty without turning vague. They let a technical argument develop, then press on the assumptions underneath it.
Reasoning at the frontier
No Priors: Artificial Intelligence | Technology | Startups brings a measured, investor-meets-engineer style to the subject. Its two-host format is patient with long answers, while the questions tend to skip introductory material and head for architecture or strategy. Really Big Test-Time Compute in AI Changes Benchmarks, Safety and Research with OpenAI Research Scientist Noam Brown is the clearest starting point here. The title alone puts three consequences of additional inference-time computation on the table: benchmarks, safety, and research.
The show also approaches capability through less conventional tests. What Chess.com Teaches US About Superhuman Capabilities, with CEO Erik Allebest offers a route into the question through chess rather than another model-release discussion. That makes No Priors useful for listeners who want reasoning placed inside a larger technology picture.
Engineers thinking out loud
Latent Space: The AI Engineer Podcast is brisker and messier in a productive way. Its interviews sound like technically fluent people working toward a sharper formulation in real time. Dense vocabulary is assumed, but the hosts keep pulling abstract points back toward mechanisms and developer consequences.
Its reasoning beat stretches from evaluation in Reality: The Final Eval — Lukas Petersson and Axel Backlund of Andon Labs to the data behind learned behavior in dY"ªCausal Models Need Causal Data - Xaira’s X-Cell model for Drug Discovery (Bo Wang & Ci Chu, Chief Discovery Officer & Chief AI Scientist). The Professor of Outputmaxxing — Anjney Midha, AMP takes yet another angle. The show rewards anyone who enjoys hearing a technical thesis assembled through questions, reframing, and the occasional useful digression.
From model behavior to stronger claims
The TWIML AI Podcast (formerly This Week in Machine Learning & Artificial Intelligence) offers direct entry points with unusually explicit titles. AI Trends 2026: OpenClaw Agents, Reasoning LLMs, and More with Sebastian Raschka places reasoning models among the year’s wider technical trends. Narrower still is Building an AI Mathematician with Carina Hong, which stays with mathematical capability. For a perceptual counterexample, Why Vision Language Models Ignore What They See with Munawar Hayat asks why apparently capable systems can fail at using available visual information.
"The Cognitive Revolution" | AI Builders, Researchers, and Live Player Analysis pushes toward recursive improvement, oversight, and model design. It's Crunch Time: Ajeya Cotra on RSI & AI-Powered AI Safety Work, from the 80,000 Hours Podcast connects recursive self-improvement with safety work. Don't Fight Backprop: Goodfire's Vision for Intentional Design, w/ Dan Balsam & Tom McGrath points toward the design choices beneath model behavior.
For a faster, more opinionated register, Machine Learning Street Talk treats AI developments as material for a running thesis rather than a formal lesson. Anthropic's Release of Opus 4.7 Overview and Meta's Gemini 4 and Future Directions show its model-by-model approach. If you want technical news translated quickly into practical consequences, with skepticism left in the mix, this is the one.
Finding the adjacent conversation
Step back from reasoning and survey the field as a whole on the Artificial Intelligence page. Frontier Models follows the systems and releases themselves, while AI Agents concentrates on models acting through tools and workflows. Alignment & Safety carries the risk and oversight questions further than this page does. AI Infrastructure is where to turn once inference hardware and deployment matter more than how a model reaches an answer.
Where to start
Begin with Really Big Test-Time Compute in AI Changes Benchmarks, Safety and Research with OpenAI Research Scientist Noam Brown for a focused frontier-research conversation. Choose Reality: The Final Eval — Lukas Petersson and Axel Backlund of Andon Labs for an engineer-facing route into evaluation. Try Building an AI Mathematician with Carina Hong when mathematical reasoning is the draw. For the safety edge, start with It's Crunch Time: Ajeya Cotra on RSI & AI-Powered AI Safety Work, from the 80,000 Hours Podcast.
Featured shows below are part of the PlayNext catalog. Updated weekly.
Featured shows
5 shows, described rather than listed
Each one comes with a short write-up and the role it plays in this guide.
The longest-form interview slot for the people designing reasoning models. Noam Brown on test-time compute, Nathan Lambert on RLVR, Ashvin Nair on the year-end o3/Cursor Composer synthesis.
Nathan Labenz's show does the most concentrated work on whether reasoning models are really thinking. Strong takes from both the bear and bull camps on mathematical superintelligence and capabilities.
Sarah Guo and Elad Gil's investor lens on reasoning models. Jensen Huang, Andrej Karpathy, and other operators read the state of the art.
Sam Charrington's long-running ML interview show. The academic-to-practitioner bridge for reasoning-era research, with episodes from Raschka and Chowdhery.
Tim Scarfe's deep methodology-first interviews. The show to listen to when you want the architecture questions pressed harder than the demos.
Where to start
3 routes in, one per reader
Start with Scaling Test Time Compute to Multi-Agent Civilizations — Noam Brown, OpenAI. The closest first-principles guide to why test-time compute works.
Is AI Stalling Out? Cutting Through Capabilities Confusion — the bear case argued cleanly, paired with Harmonic's mathematical-superintelligence episode for the bull case.
Jensen Huang's episode is the status read; Andrej Karpathy's is the closest read on what's actually new about the reasoning era.
Glossary
Terms worth knowing before you listen
Compute spent at inference time on iterative reasoning, search, and self-verification, rather than at training time. The core architectural shift behind o1 / o3 / R1-class models.
Training reasoning models on tasks with checkable answers (math, code, formal proofs) where the reward signal is unambiguous. Nathan Lambert's preferred frame for the o-series recipe.
An intermediate-reasoning trace produced by the model before its final answer. The simplest form of test-time-compute scaffolding; later reasoning models internalize and extend it.
International Mathematical Olympiad gold-medal performance — the canonical 2025 benchmark for reasoning models. Harmonic, OpenAI, and DeepMind have all claimed variants.
Mechanistic-interpretability angle — Goodfire's argument that intentional design of internal model structure may be the next reasoning-gains frontier beyond pure RL post-training.