Most AI podcasts stop to explain what a neural network is, and if you write code, you skip them. These six do not. On one, the hosts drop the guest and just argue the week's news. On another, a host takes AlphaFold apart with the Nobel winner who built it. Shows that only book guests with a model to sell did not make the list. Some run solo; some sit inside networks like Changelog and Turpentine. These are the AI podcasts for engineers who actually build things. A player sits under each pick.
1. Latent Space: built for AI engineers, with an essay for every episode
swyx and Alessio Fanelli both write code for a living, and that shapes who they invite on. They launched Latent Space in February 2023 as a show made by and for AI engineers. Every weekly episode comes with a written newsletter, which reached more than ten million readers in 2025. Andrej Karpathy has been a guest, along with Simon Willison and Modal's Erik Bernhardsson. The bar is high. Because both hosts have shipped real software, the guests tend to be builders rather than press reps. They also run the yearly AI Engineer conference. You can read the essay or just press play.
The guests are the reason to stay, and they range from lab founders to solo open-source hackers. One week it might be a Databricks cofounder explaining how databases change for AI agents. The next, swyx sits alone and maps where the whole field is heading. Those solo recaps are quietly the best part. Newcomers can feel lost, because the hosts assume you already read last week's paper. That is the price of a show built for working engineers, and it is worth paying. Building with models right now? Start here. For pure research over engineering, try the Dwarkesh Podcast in section six.
2. Machine Learning Street Talk: edited like a documentary, not an interview
MLST plays more like a documentary than an interview. Tim Scarfe started it in April 2020, and MIT's Keith Duggar joins him for most episodes. Each one opens cold, then walks you through a model's inner workings with the researcher who built it. When Nobel laureate John Jumper came on, they took AlphaFold apart piece by piece. It runs weekly. Michael Jordan, Beth Barnes and Taco Cohen have all faced the same careful questioning. The tagline promises hype surgically removed, and the editing lives up to it. Expect deep talk on the science and philosophy behind AI, with nothing dumbed down.
Everyone argues here, and that back-and-forth is the engine of the show. Scarfe keeps pressing until a guest defends the actual idea, not the marketing around it, and Duggar pushes back too. Rigor costs time. A single conversation can run two hours and wander into some very abstract math. If you just want a headline, this is the wrong show. But if you want to understand why a model really works, almost nothing else goes deeper. You spend the minutes, and you walk away with the mechanics. It suits working researchers, graduate students, and anyone who reads papers for fun.
3. The TWIML AI Podcast: one analyst, numbered episodes, since 2016
Since 2016, Sam Charrington has hosted this show alone, and the preparation shows. As an industry analyst, he sits with one guest at a time and goes deep. Every episode is numbered and comes with a detailed notes page on his site. By mid 2026 he had passed 770 of them. The range is wide. One week Alex Wiltschko explains how you teach a computer to smell. The next, Stanford's Stefano Ermon walks through a new kind of language model. Sebastian Raschka has stopped by to sum up the year in LLMs. The tone stays calm and technical, never loud or jokey.
Because he does the reading, his questions land exactly where the work is. The trade-off is energy. Nobody here is cracking jokes, and the pace stays even and grown-up. If you want spark, it will feel flat, but that calm is what lets a guest go properly technical. Guests get room to think, and the silences are allowed to sit. For something more hands-on, Practical AI in section four is close by. Nine years of numbered episodes add up to an archive deep enough to teach you a whole subfield. Start with a guest whose field you already know, and then follow the numbers back.
4. Practical AI: the hosts drop the guest to catch you up
Some weeks there is no guest at all. Daniel Whitenack and Chris Benson call those Fully Connected episodes, where the two just talk through the week's news. No guest, no pitch. It works surprisingly well. Practical AI has run on the Changelog network every week since July 2018. When they do bring someone on, it is usually a builder. Nous Research's Jeffrey Quesnelle came on to talk about agents, and Comma AI's Harald Schäfer explained open-source self-driving. Even Congressman Don Beyer, who is studying for an AI PhD, has appeared. The show keeps its focus on shipping code, not theory.
You come back for how easy it is to listen to. These are two friends who know AI well and never talk down to you. In one Fully Connected episode, they walked through the annual Stanford AI Index together. The catch is depth. They cover a lot of ground but rarely go all the way down, so specialists may want more. In return you get a show that fits neatly into a commute. It also welcomes practitioner guests, so save it for the drive to work, when you want company more than a lecture.
5. The Cognitive Revolution: hosted by one of GPT-4's pre-launch red teamers
Nathan Labenz got to test GPT-4 months before the public did. He was on the team that red-teamed it before launch, and he is named in the model's technical report. That is rare access. He and Erik Torenberg started The Cognitive Revolution in early 2023, on the Turpentine network. Labenz tries the newest models himself before he interviews anyone about them. His guests have included David Dalrymple on AI safety and Kunle Olukotun on the chips underneath it all. New episodes land close to weekly, and they run long. Together the two of them track the frontier as it keeps moving.
Length is the cost. Episodes often run past two hours, with frequent detours into safety and policy. Because Labenz has actually used these tools under pressure, he asks what breaks, not just what impresses. If you want a quick summary, look elsewhere. But when you need a clear picture of what AI can and cannot yet do, few hosts are better prepared. He also posts hosts-only recap episodes between the big interviews. Those long hours buy you a genuine read on where things really stand. For similar depth with more of the underlying math, Machine Learning Street Talk in section two is the pick.
6. Dwarkesh Podcast: which AI podcast preps hardest before the mic?
Nobody on this list prepares like Dwarkesh Patel. Before each interview he reads the relevant books and papers, then uses AI tools he built himself to organise them. He has said Claude and Cursor do a lot of that sorting. All that homework means he can challenge a guest inside their own field. Jensen Huang has sat down with him, as have Dario Amodei and the analyst Dylan Patel. Even Tony Blair and Mark Zuckerberg have taken a turn. He has run the show since 2020, weekly, with episodes that stretch two to three hours. Prep is the point. You hear frontier ideas straight from the people building them.
The tone mixes genuine respect with steady pushback. Patel never flatters. He will question a guest's timeline or scaling story the moment they say it. One episode dug into why the search method behind AlphaGo does not carry over to language models. Another toured a Jane Street data centre with the floorboards literally pulled up. The catch is focus. The show roams across AI, economics, history and biology, so it reads more like a worldview than a how-to. It will not teach you to ship code, but for thinking clearly about where AI is heading, nothing beats it. Set aside an afternoon; these run long and earn it.
So where do you start? If you build with models daily, open Latent Space and Practical AI first. If you want the frontier instead, Dwarkesh and MLST reward the hours. Pick one, not five. Play a single episode on your next commute, and see if the host earns your time before you subscribe. The rest of the field can wait until next week.