The best AI podcasts, sorted by who each show is actually for

Search for the best AI podcast and you’ll get twenty lists ranking roughly the same ten shows, almost none of which say what any show is actually for. But that’s the wrong question to start from. A daily news brief and a two hour research interview aren’t competing with each other, because they answer different needs: keeping up, understanding a method, or working out where the money is going. So this piece sorts the shows by the job you want done, and then checks each one against its own RSS feed rather than against its marketing.

We pulled every feed below on 21 August 2026 and measured three things from the XML: how many episodes the show published in the previous 90 days, the median length of its most recent 26 episodes, and the minutes a week those two numbers imply. The short version, if you want it now: The AI Daily Brief for the news, TWIML and Machine Learning Street Talk for the research, Latent Space and Practical AI if you’re building, No Priors and Dwarkesh for the money and the arguments. So the rest of this explains what each one costs you in hours, and where the feeds disagree with the reputations.

The audience for AI podcasts is the audience for AI

Start with why this category exists at all, because the overlap is unusually tight. Edison Research at SSRS published The Infinite Dial 2026 on 12 March 2026, from a January 2026 national survey of 2,050 Americans aged 12 and over. It found that 45% of Americans aged 12 and over, about 130 million people, had consumed a podcast in the last week. That same study added generative AI questions for the first time, which is what makes it useful here.

The finding that matters here is the crossover. Edison reports that “more than half of AI users are weekly podcast consumers, versus roughly one-third of AI non-users”, and that 87% of AI users listened to online audio in the last week against 61% of non-users. That’s why every lab, fund and developer tools company now has a show. The audience they want is already wearing headphones.

57% of Americans age 12+ have ever used a generative AI assistant, a milestone that took podcasting 16 years to reach.

The Infinite Dial 2026, Edison Research at SSRS, March 2026

Which means the supply side is crowded, and the crowding is the problem you’re actually solving. There are more good AI shows than anyone has hours for, so picking well is a subtraction exercise. The table below is the shape of the answer, and the sections after it explain the calls.

If you want	The show	Format
The week’s news, compressed	The AI Daily Brief, Last Week in AI, Hard Fork	Monologue or two-host roundup
Researchers explaining their own work	TWIML, Machine Learning Street Talk	Long technical interview
A lab’s own account of a result	Google DeepMind: The Podcast	Produced, presenter-led
To ship something this quarter	Latent Space, Practical AI	Practitioner interview
To understand where capital is moving	No Priors, Dwarkesh Podcast	Investor and founder interview

If you want the week’s news in the time it takes to make coffee

The AI Daily Brief is the workhorse here, and the feed backs up the name. It published 82 episodes in the 90 days to 21 August 2026, with a median length of 29 minutes across its last 26. Nathaniel Whittemore hosts it solo, and the show describes itself as “a daily news analysis show on all things artificial intelligence”. Recent episodes include “The AI Backlash Is Getting Stupider. But Also Smarter.” and “AI Companies Still Haven’t Delivered on Their Biggest Promises”, which tells you the register: opinionated synthesis, not a wire feed.

The catch is volume. At that cadence you’re being offered roughly three hours a week, which is more than most people will give a briefing. Treating it as a buffet rather than a subscription works better, though, because the episode titles are descriptive enough to make skipping easy.

Last Week in AI is the calmer alternative, run by Andrey Kurenkov as a paired newsletter and show, described in its feed as “weekly text and audio summaries of the most interesting AI news, as well as editorials commenting on recent events”. Episode #253 covered Opus 5, Gemini 3.6, Kimi K3 and a Hugging Face hack. Its weakness shows up in the feed itself, though: the show hadn’t published since 3 August as of 21 August, and six of the episodes in its feed carry the same 21 July publication date, so the dates can’t be read as a clean cadence.

Hard Fork is the third option and the least technical of the three. It’s a New York Times show, and the paper describes it as journalists Kevin Roose and Casey Newton making sense of the rapidly changing world of tech. Fourteen episodes in 90 days at a median of 66 minutes, and the titles show three segments in almost every one. If you already read the news, it’s the one that adds the least, which isn’t a criticism so much as a description of who it’s built for. We’ve written separately about building a reading stack that survives the volume, and a weekly show like this one sits comfortably at the top of it.

If you want researchers explaining their own work

Sam Charrington’s TWIML AI Podcast is the longest-running show on this list, with 791 episodes in its feed going back to May 2016. The format hasn’t changed much either, which is part of the appeal: one researcher, one piece of work, about an hour. Episode #771 was “How AI Learns to Smell with Alex Wiltschko”, and #770 was “Why AI Agents Break the GenAI Security Model with Devvret Rishi”. The show’s own description says it brings “the top minds and ideas from the world of ML and AI” to researchers, data scientists and engineers, and it’s honest about that audience.

The thing the feed reveals is a slowdown. TWIML published five episodes in the 90 days to 21 August 2026, which is roughly one a fortnight for a show whose original name was This Week in Machine Learning and Artificial Intelligence. Still, that’s 22 minutes a week of high-density material, so it isn’t a reason to skip it. It does mean it can’t be your news source.

Machine Learning Street Talk sits next to it and goes harder. Its feed promises discussions covering “current affairs in AI, cognitive science, neuroscience and philosophy of mind with in-depth analysis”, and the median of its last 26 episodes is 78 minutes. Its recent run included an interview with Matthieu Wyart on the level of abstraction models learn at, and one recorded with Apollo Research on how researchers test AI for hidden goals. It’s the show most likely to spend an hour on a disagreement rather than resolving one, so it rewards patience and punishes half-attention.

Then there’s the lab’s own show. Google DeepMind: The Podcast, presented by the mathematician Hannah Fry, promises in its feed to go behind the scenes of the lab with “no hype, no spin”. Recent episodes are titled “Understanding the inner thoughts of AI” and “When millions of AI agents meet”. It’s beautifully produced and it’s also corporate media, so the useful posture is to treat it as a primary source about what DeepMind wants understood, rather than as journalism. But the bigger practical limit is supply: two episodes in 90 days, and nothing since 10 July 2026.

If you’re the one shipping the thing

Latent Space is the show for people whose job is to put models into production. It calls itself “the podcast by and for AI Engineers” and says that in 2025 over 10 million readers and listeners came to it, which is the show’s own claim rather than an audited figure. It published 19 episodes in the last 90 days at a median of 76 minutes, which is the second highest output on this list. Recent guests came from Baseten on inference engineering, from OpenAI on scaling Codex, and from Poolside on building a model factory.

That guest list is also the thing to watch. Episodes are frequently built around a company explaining its own product, which makes them useful and interested at the same time. The specifics tend to survive the sales pitch, because the details of a serving stack are hard to fake in ninety minutes.

Practical AI is the gentler entry point, hosted by Daniel Whitenack and Chris Benson, with the stated aim of making artificial intelligence practical, productive and accessible to everyone. It runs roughly weekly at a median of 47 minutes, and the recent run included “Models, Harnesses, and Multi-Agent Systems” and “Reconstructing how OpenAI agents attacked Hugging Face”. Latent Space assumes you already have the vocabulary, and Practical AI assumes you’d like to be told. Our own glossary of what the words mean when a vendor uses them covers the same gap in text.

If you’re deciding where the money goes

No Priors is co-hosted by Sarah Guo, founder of the investment firm Conviction, and Elad Gil, described on the show’s own listing as a serial entrepreneur and startup investor. Its feed says the co-hosts talk to “the world’s leading AI engineers, researchers and founders” about how far away AGI is and which markets are at risk. That works out at twelve episodes in 90 days and a median of 41 minutes, which makes it the easiest weekly commitment in this group. August 2026 episodes covered restoring sight and reimagining the brain with Max Hodak, and superhuman capability with the chief executive of Chess.com.

The conflict is structural rather than hidden. An investment firm’s podcast interviews founders, and investment firms back founders, so you’re hearing a curated slice of the market from people with positions in it. That’s still worth your time, as long as you read it as deal flow rather than analysis. Our piece on how AI companies actually make money is a useful counterweight when an episode gets enthusiastic about a business model.

Dwarkesh Patel’s show is the other half of this category, and it asks more of you than any other dedicated AI show here. The median of its last 26 episodes is 101 minutes, and it bills itself simply as “deeply researched interviews”. Recent output included an interview with Ryan Greenblatt on what happens once AI can automate AI research, and a piece arguing that smarter models could push compute prices up tenfold. The preparation shows, which is the reason to pick it, but the length is the reason people don’t finish it.

What the feeds say about the time each show wants

Reputation and output diverge more than the recommendation lists admit, which is easier to see in one place. Here is what the XML says, measured the same way for every show.

Show	Publisher or host	Episodes, 90 days	Median length	Minutes a week
The AI Daily Brief	Nathaniel Whittemore	82	29 min	185
Latent Space	Latent.Space	19	76 min	112
Dwarkesh Podcast	Dwarkesh Patel	10	101 min	79
Hard Fork	The New York Times	14	66 min	72
Lex Fridman Podcast	Lex Fridman	4	208 min	65
Machine Learning Street Talk	MLST	8	78 min	49
No Priors	Conviction	12	41 min	38
Practical AI	Whitenack and Benson	10	47 min	37
TWIML AI Podcast	Sam Charrington	5	57 min	22
Google DeepMind: The Podcast	Google DeepMind	2	46 min	7
Last Week in AI	Andrey Kurenkov	9	not published	not calculable

Two things jump out of that column on the right, though. The AI Daily Brief alone asks for more of your week than any three of the research shows combined, and the lab-branded show and the longest-running one, DeepMind and TWIML, are the two quietest on the table. If you’ve got about two hours a week, one research show plus one weekly roundup fits inside it, because those two columns add up to roughly 120 minutes. The daily brief on its own does not.

The show everyone recommends that mostly isn’t about AI

The Lex Fridman Podcast appears near the top of almost every AI podcast list, and its feed makes that hard to defend in 2026. It published 23 episodes in the year to 21 August 2026. Three of them name AI, an AI agent or an AI company in the title: Jensen Huang of Nvidia in March, an episode on the OpenClaw agent in February, and a state of AI roundup on 1 February.

The other twenty covered the Roman Empire, Vikings, nuclear fusion, Khabib Nurmagomedov, the Civil War, FFmpeg and dark energy. That’s consistent with the show’s own description, which lists AI as one of about twenty subjects. So it isn’t a bait and switch, but it does mean that recommending it as an AI podcast is a category error, and the episodes run a median of 208 minutes. It belongs on your list as a long-form interview show that occasionally covers AI.

The same check is worth running on any show you’re about to subscribe to, because reputations lag feeds by a year or more. Open the feed, read the last twenty titles, and see whether the show is still doing what its reputation says.

What a podcast can’t do for you

Give the sceptical case its best version, because it’s a strong one. Audio has no correction mechanism worth the name. A written post gets a dated note appended when a number turns out to be wrong, but an episode published in June stays wrong in your app forever. That’s the single biggest reason not to treat any of these shows as your source of record.

Benchmark claims are where this bites hardest. A guest quoting a score on air gives you no way to check the evaluation setup, the contamination controls or the version of the model tested, and those are the details that decide whether the number means anything. We’ve laid out why benchmark numbers keep misleading people at length, and the short version is that a figure without its methodology is decoration.

Then there’s decay. Model releases have been arriving every three weeks or so, which means a technical episode about serving costs or context limits can be out of date within a quarter, and nothing in your podcast app tells you which ones have expired. The interview format also rewards confidence over calibration, because a guest who says “we don’t know” makes for worse audio than one who doesn’t. That’s not dishonesty on anyone’s part, it’s the medium doing what the medium does.

What podcasts do well, better than text, is expose reasoning. You hear a researcher pause, qualify, and disagree with the host, and you learn how the person thinks about the problem. That’s worth an hour when the alternative is a press release. It just isn’t a substitute for reading the paper, which is also why we’d point anyone using these shows for research towards checking sources rather than trusting a summary.

What would change this list

Three things, and they’re all measurable from the feeds. If TWIML’s cadence returns to something close to weekly, it becomes the default recommendation for the research category rather than a shared one, because its back catalogue is unmatched. If DeepMind’s show resumes at the pace it kept through late 2025, when it published six episodes between October and December, the lab category stops being a footnote.

The third is the one to watch hardest. Edison found that 57% of Americans over 12 have used a generative AI assistant, which took podcasting 16 years to match. A market that size attracts sponsors, so sponsor money changes what gets made. If the shows in the research column start looking like the shows in the investor column, the useful split this piece is built on stops holding, and you’d want to rebuild the list from the feeds again.

Until then, though, the method matters more than the ranking. Pick the job first, check the feed second, and let the minutes-a-week column decide what you actually subscribe to.