Topics / Frontier Models
Frontier Models Podcasts
10 shows and 5,052 episodes sit under Frontier Models.
How the top AI podcasts are framing frontier-model capability, evaluation, and the open-weight gap — survey of recent episodes from No Priors, Daily Paper Cast, Brain Inspired
Photo: Brecht Corbeel on Unsplash · Unsplash License
Swipe · Frontier Models only
Swipe through Frontier Models
Save time discovering your next favourite podcast by listening to just the best minute of an episode.
The frontier models slate · every cover a link
All 10 frontier models shows →
Next to this on the shelf
Pick a subtopic
12 subtopics
Your guide
How Frontier Models podcasts differ, and which to try
The release itself is usually the least interesting part. The useful listening begins after the benchmark graph, when someone asks what produced the gain, whether it survives contact with a real task, and who can afford to repeat it. Frontier-model podcasts range from research briefings dense with inference jargon to reporting on the companies turning capability into power.
For the wider field, move up to Artificial Intelligence. AI Reasoning narrows the focus to how models solve problems, AI Infrastructure follows the compute underneath them, and Alignment & Safety concentrates on control and risk. This page stays with the models pushing the boundary and the labs trying to build them.
The labs as a contest
FT Tech Tonic brings editorial restraint to a subject that often attracts breathless commentary. Its lab series establishes the competitive field with The battle of the AI labs, then gives individual companies sustained attention in AI Labs: Google DeepMind plans its comeback, AI Labs: Sam Altman may make or break OpenAI, and AI Labs: Are Anthropic really the good guys? For reporting on the race rather than cheering from the sidelines, its measured narration and scene-led production make it a strong choice.
Decoder with Nilay Patel is sharper and more argumentative. The host’s long, nested questions test a guest’s premises, while his reductive restatements expose the business decision hiding inside the technical language. Does Google even want to win in AI? puts strategy ahead of benchmark watching; Welcome to the AI crisis in math turns toward capability; and The left and right agree on one thing: no data centers follows the physical resistance surrounding the buildout. It is especially good when a company’s public story sounds tidier than its incentives.
Capability, close up
No Priors: Artificial Intelligence | Technology | Startups assumes the listener can keep up. Its two hosts work like informed peers, moving quickly past introductions and into architecture, strategy, and tradeoffs. Really Big Test-Time Compute in AI Changes Benchmarks, Safety and Research with OpenAI Research Scientist Noam Brown is the clearest technical on-ramp, while Chasing Trillion-Dollar Companies, Founder Ambition, Token Budgets, and Regulatory Capture with Sarah & Elad widens the frame around the engineering. The result feels closer to a serious industry briefing than a product launch recap.
Daily Paper Cast operates at research-paper speed. Its synthetic co-hosts move through methods and metrics in an orderly question-answer rhythm, with minimal digression and little fear of specialist vocabulary. Self-OPD: On-Policy Distillation for Flow Matching Models without Teacher and TTPO: Test-Time Policy Optimization go straight to training and inference techniques; PAWBench: How Far Are We from Probabilistically Aligned World Modeling? and FrontierChallenge: Evaluating Scientific Workflow Completion concentrate on evaluation. This is the efficient option when you would rather hear a paper unpacked than listen to an hour of industry scene-setting.
The system around the model
Lex Fridman Podcast gives frontier AI room to sprawl. The long-form interview style is reverent, patient, and willing to wander into philosophy before returning to technical ground. #490 – State of AI in 2026: LLMs, Coding, Scaling Laws, China, Agents, GPUs, AGI gathers the major fault lines into one conversation, while #494 – Jensen Huang: NVIDIA – The $4 Trillion Company & the AI Revolution shifts attention toward the company most closely associated with the compute boom. Choose it when duration is an advantage and you want the surrounding worldview left intact.
Daily Tech News Show offers the opposite tempo: brisk conversation, dry asides, and a steady return to the day’s headline. OpenAI Slows Down and Focuses on Safety - DTNS 5335 tracks a change in emphasis at one lab, while Anthropic is (almost) Not a Supply Chain Risk - DTNS 5342 and NVIDIA is about to buy HuggingFace - DTNS 5341 connect model development to vendors and platforms. It works best as a running pulse check between deeper interviews.
Where to start
Begin with FT Tech Tonic’s The battle of the AI labs for the competitive landscape. Follow it with No Priors: Artificial Intelligence | Technology | Startups and Really Big Test-Time Compute in AI Changes Benchmarks, Safety and Research with OpenAI Research Scientist Noam Brown for the technical pressure point. Daily Paper Cast’s FrontierChallenge: Evaluating Scientific Workflow Completion is the compact research choice; for one expansive overview, take Lex Fridman Podcast and #490 – State of AI in 2026: LLMs, Coding, Scaling Laws, China, Agents, GPUs, AGI.
Featured shows below are part of the PlayNext catalog. Updated weekly.
Featured shows
10 shows, described rather than listed
A short write-up for each, plus the job it does in this guide.
Sarah Guo and Elad Gil host the operator's frontier — chips, clouds, agentic engineering, capital. The Karpathy 'Loopy Era' episode (March 2026) is the slate's reference point for what current frontier models actually do once you put them in a loop.
Short, dense episodes that surface new model and alignment research without packaging. The closest the slate gets to research-as-it-ships rather than research-once-it's-a-press-release.
Paul Middlebrooks runs three-hour conversations with researchers studying how brains and learning machines diverge. Listen before any earnings-call sentence about 'reasoning' or 'understanding.'
The longest-form interview show with the principals running the labs — episodes #494 (Jensen Huang) and #490 ('State of AI in 2026') function as primary source material rather than as guides. High signal, but you have to do the editorial work yourself.
Nilay Patel takes the AI strategy questions other shows compress into five minutes and runs them across an hour. The frontier-model coverage shows up as platform-and-product reasoning rather than benchmark commentary.
Tom Merritt's daily news read is where frontier-model news lands first — Anthropic-Pentagon, OpenAI funding, lab-state alignment. Useful as velocity tracking; pair with one of the depth shows.
Daniel Whitenack and Chris Benson cover the application layer once a frontier capability is stable enough to deploy. 'Are we in an AI bubble?' and 'How is AI shaping democracy?' are representative recent entries.
The Financial Times' magazine-format AI reporting. 'Tech in 2026: Inside the AI bubble' is the closest recent episode to a sustained narrative read on the frontier-model conversation.
Kyle Polich's long-running show on algorithmic critique. Recent episodes lean heavily on recommender-system fairness and interpretability — the literature most relevant to 'what does an improvement actually measure.'
Adam Stacoviak and Jerod Santo cover the open-source AI ecosystem with sharper skepticism than the VC-funded shows. 'A new direction for AI developer tooling' is a recent representative entry.
Where to start
4 routes in, one per reader
Karpathy's 'Loopy Era' episode and the recent agentic-engineering installments treat frontier capability as an engineering surface, not a magic trick.
Daily news episodes are where Anthropic-OpenAI-Google announcements land first; faster than the depth shows, lower context.
Short episodes per paper, no editorializing — useful as a weekly research check-in.
#494 with Jensen Huang and #490 'State of AI in 2026' are primary-source listens. Three-plus hours, philosophical drift permitted.
Glossary
A quick Frontier Models glossary
Informal name for the largest, most capable AI models at any given moment — currently the GPT, Claude, and Gemini families, with the open-weight contingent (Llama, Qwen, Mistral) closing distance on different axes than the leaderboards measure.
An AI model whose trained parameters are publicly downloadable, even if the training data and code aren't. Distinct from 'open source.' The closed-vs-open question is one of the most consequential structural questions in AI right now.
What happens when a public evaluation set leaks into training data. The model 'improves' on the benchmark without actually getting better at the underlying task. Why simple leaderboard-watching is a noisier signal than it looks.
Andrej Karpathy's framing on No Priors (March 2026) for the current state of agents — calling tools, calling themselves, accumulating mistakes that mostly compound and occasionally don't. The phrase has shaped subsequent episodes.