The 11 Best AI Engineering Podcasts in 2026, Benchmarked
Also published in 
Medium (opens in new tab)
Every “best AI podcasts” list says the same thing. The same shows, the same recycled blurbs, and no evidence the writer listened to a single episode. Worse, none of them give you a way to check the claims. So I benchmarked them. The same way you’d benchmark a model before you trust it, on July 5, 2026 I pulled public catalog and feed data on eleven leading AI engineering podcasts: catalog depth, publishing cadence over the trailing six months, average episode length, and whether the show ships machine-readable transcripts. Every number here is reproducible from a public RSS feed or Apple’s podcast lookup API.
The short version: eleven shows, benchmarked on public feed data. Only three ship machine-readable transcripts, only two clear 90%, and exactly one exposes its download numbers.
Full disclosure: I host one of the shows below, Chain of Thought. That’s exactly why the list is alphabetical instead of ranked, and why my show gets the same treatment as everyone else’s. Discount my bias accordingly. And yes, the benchmark surfaces some unflattering facts about my own show too. You’ll see.
Two findings before the list. First, the cadence spread is enormous: the most prolific show benchmarked publishes more than eight times as often as the least. Second, transcripts are still rare. Of the eleven shows, only three include machine-readable transcripts in their RSS feed at all, and only two cover more than 90% of their episodes. That gap matters more than it used to: in 2026, search engines and AI assistants increasingly surface podcasts by reading their transcripts, not their audio. A show with no transcript is invisible to them.
The data
All eleven shows are below. Figures are as of July 5, 2026; cadence, length, and transcript coverage are computed over the trailing six months.
| Show | Catalog | Episodes/month | Avg length | Transcripts | Downloads |
|---|---|---|---|---|---|
| AI Engineering Podcast | 79 | 1.0 | 55 min | 83.3% | Public (OP3) |
| Chain of Thought | 66 | 3.2 | 53 min | 100% | Private |
| The Cognitive Revolution | 357 | 8.3 | 117 min | 0% | Private |
| Dwarkesh Podcast | 132 | 2.8 | 103 min | 0% | Private |
| Gradient Dissent | 138 | 1.2 | 59 min | 0% | Private |
| How I AI | 89 | 7.5 | 42 min | 0% | Private |
| Latent Space | 212 | 7.8 | 72 min | 0% | Private |
| MLOps.community | 535 | 7.5 | 55 min | 0% | Private |
| No Priors | 169 | 4.0 | 43 min | 0% | Private |
| Practical AI | 364 | 3.5 | 48 min | 100% | Private |
| The TWIML AI Podcast | 788 | 1.8 | 63 min | 0% | Private |
A note on the transcript column: it measures transcript tags in the show’s public RSS feed, the machine-readable kind podcast apps and search engines can use. Some shows, Latent Space in particular, publish written editions on their websites instead, which is great for human readers but invisible to most podcast apps.
That last column is partly a confession. Download numbers stay private unless a show routes its feed through OP3, the open podcast analytics project, which makes them public. Exactly one of the eleven does: the AI Engineering Podcast. Everyone else, my own show included, keeps download stats to ourselves. So treat every “most popular AI podcast” claim you read, including any I might make, with the skepticism it earns. If you want to read OP3 data, for that show or any other that adopts it, I built op3-mcp to pull it straight from an agent.
The shows
AI Engineering Podcast
The AI Engineering Podcast is Tobias Macey’s interview show about the operational side of AI: infrastructure, deployment, governance, the unglamorous work that makes models useful in production. Its feed has gone quiet, though. The most recent episode, “Kubernetes, Compliance, and Control: The Operational Backbone of AI Sovereignty,” landed in late February 2026, so the cadence figure reflects a burst early in the window rather than a steady monthly clip. It’s also the one show here whose downloads are actually public, because it routes its feed through OP3, and transcript coverage over the window runs 83.3%, five of its last six episodes. The 79-episode back catalog is the real draw, and it goes deep on a single system or practice each time.
Best for: platform and infrastructure engineers who own AI systems in production.
Chain of Thought
My show, so judge this entry hardest. Chain of Thought is practitioner interviews with the engineers and leaders building production AI systems, with guests from NVIDIA, Google, Databricks, and Cohere. The benchmark: 66 episodes, still the youngest catalog in the set, a 53-minute average length, and 100% transcript coverage in the feed, tied with Practical AI for the highest here. What it won’t show you is downloads: like almost everyone on this list, I don’t route through OP3. A recent episode that shows the format: Jerry Liu of LlamaIndex arguing that the AI framework era is over and context is the moat.
Best for: AI engineers and engineering leaders who want operator-level detail without a 90-minute commitment.
The Cognitive Revolution
The Cognitive Revolution is the most demanding show on this list on both axes: the highest cadence in the set at 8.3 episodes a month, and the longest average at 117 minutes, across 357 episodes. Nathan Labenz and Erik Torenberg host it, and Labenz approaches AI like a scout, testing tools and interrogating builders about what’s changing. A recent episode with Liquid AI’s Ramin Hasani dug into device-native foundation models and pushing intelligence to the edge. The volume means uneven peaks, but the peaks are high.
Best for: people who treat AI progress as a beat to follow daily and have the commute to match.
Dwarkesh Podcast
Dwarkesh Patel’s show does long-form interviews with researchers, economists, and lab leaders, and he prepares harder than almost anyone in the medium. The benchmark backs the reputation for depth: 132 episodes at a 103-minute average, among the longest of the set, published a couple times a month. The questions assume you’ve done the reading. A recent episode with Grant Sanderson on AI and the future of math is the kind of deep, unhurried conversation this show exists for. Full transcripts live on his site, though not as tags in the feed, so they don’t count in the column above.
Best for: listeners who want the research and macro picture behind the engineering.
Gradient Dissent
Gradient Dissent is Lukas Biewald, CEO of Weights & Biases, interviewing founders and researchers at the frontier of applied ML. The cadence is one of the slowest in the set, about an episode a month, but the guest list compensates: a recent episode featured Dan Klein on building an AI that can’t lie. Episodes average 59 minutes, and the 138-episode catalog doubles as a history of the modern ML tooling industry, told by the people who built it.
Best for: ML practitioners and founders who want the company-building view.
How I AI
How I AI is Claire Vo’s show about how people actually use AI tools to get work done, and it’s a personal favorite of mine. Less about model internals, more about the workflow: what someone opens, prompts, and ships in a real day. The benchmark puts it at 89 episodes, a brisk 7.5 per month, and the tightest average in the set at 42 minutes, with no transcripts in the feed yet. A recent episode ran 64 generations to stress-test whether Sonnet 5 is worth it, which is exactly the hands-on register the show works in. It’s also on YouTube.
Best for: builders who want concrete, over-the-shoulder AI workflows instead of theory.
Latent Space
Latent Space is Swyx and Alessio Fanelli running what amounts to the trade journal of AI engineering: 7.8 episodes a month over the trailing six months, one of the highest cadences in the set, at a 72-minute average. The show tracks the AI engineering discipline as it forms, from evals to agents, and its conference coverage means you hear practitioners you won’t find elsewhere. A recent episode with Evan Feinberg and Sergey Edunov of Genesis Molecular AI argued the most interesting diffusion research isn’t happening in LLMs. The feed carries no transcript tags, though written editions live on the Latent Space site.
Best for: working AI engineers who want to stay current week to week.
MLOps.community podcast
Demetrios Brinkmann’s community-run show is the closest thing this field has to an open mic for production war stories: practitioners on what broke, what scaled, and what they’d do differently. It’s also the workhorse of the set, 535 episodes deep and publishing 7.5 a month at a 55-minute average. The agenda is set by a community of thousands of ML and platform engineers rather than a media company, and it shows in the topics. A recent episode, “Omnigent: Composition, Control, and Collaboration for AI Agents,” is typical of where the conversation is in 2026.
Best for: ML and platform engineers who want peer-level honesty over polish.
No Priors
No Priors is Sarah Guo and Elad Gil bringing the investor lens, and the access that comes with it. It’s one of the more time-efficient shows in the set: a 43-minute average at 4.0 episodes a month across a 169-episode catalog. The show ranges across the frontier; a recent episode featured Valar Atomics founder Isaiah Taylor on using nuclear power to unlock energy abundance. You get the strategic view from the people allocating the capital, compressed.
Best for: engineers who want the business and capital-markets context in under 45 minutes.
Practical AI
Practical AI, from Daniel Whitenack and Chris Benson, is the most approachable show on this list without dumbing anything down: 364 episodes, a steady 3.5 per month, 48-minute average. Where other shows chase frontier discourse, Practical AI stays applied, like a recent episode on image generation and visual intelligence with Black Forest Labs. It’s also one of only two shows here with near-complete transcript coverage, at 100%, tied with Chain of Thought.
Best for: developers earlier in their AI journey, and anyone who values consistency over spectacle.
The TWIML AI Podcast
The TWIML AI Podcast is Sam Charrington’s show, and it has the deepest archive in the field: 788 episodes, more than double any other here. The current cadence is a steady 1.8 episodes a month at a 63-minute average, and the format is a focused single-guest technical interview. A recent episode, “Why AI Agents Break the GenAI Security Model” with Devvret Rishi, is exactly the kind of grounded, application-specific conversation TWIML does well. If you want to trace how an idea evolved over a decade of ML, this catalog is the primary source.
Best for: ML engineers and researchers who want technical depth with a long memory.
How to actually choose
If your constraint is time, How I AI and No Priors keep it tightest, both close to 40 minutes an episode. If your constraint is depth, The Cognitive Revolution and Dwarkesh publish more hours of frontier conversation per month than most people can absorb. If you learn by searching and reading rather than listening, the transcript column is your guide, and only two shows clear 90%. And if you want the historical record, TWIML’s 788 episodes have no peer.
The bigger point: you can ask a podcast recommendation for evidence the same way you’d ask a model vendor for a benchmark. Cadence, catalog, length, and transcript coverage all sit in public feeds, free for anyone to check. I’ve shown my numbers and my method, including the two places my own show comes up short.
But picking the right show for you will mean actually listening to a couple of episodes, and deciding which one you enjoy, or find value in.
Methodology: catalog depth from Apple’s public podcast lookup API, with Podcast Index as fallback; cadence, average duration, and transcript coverage computed from each show’s live public RSS feed over a 183-day window ending July 5, 2026. Transcript coverage counts machine-readable transcript tags in the feed. OP3 availability is read from each show’s feed prefix. Download numbers and chart positions are private or ToS-restricted and were deliberately left out. The tooling is open: podcast-benchmark for the feed metrics, op3-mcp for the OP3 reads.