A podcast usually ends when someone takes off their headphones. This one, according to a listener who builds software for a living, carried on into the product. In February 2026, Shortwave cofounder Andrew Lee described hearing a Cognitive Revolution interview with the founders of Bolt. Their explanation of an AI agent prompted his team to experiment with a fully agentic approach. He reported that it worked. The review had a useful specificity: someone heard something, then tried something.
- Long AI interviews with people building and researching the systems.
- Free audio and video, plus transcripts and newsletter updates.
- A host who asks about implementation, costs, limitations and consequences.
That is a good place to begin with The Cognitive Revolution, an AI media brand launched by Nathan Labenz and Erik Torenberg in 2023. Its offering is straightforward: conversations and analysis for people trying to understand what AI can do. Its promise becomes more interesting when the audience includes someone who can take an explanation back to an engineering team. A listener becomes an experimenter. The interview has acquired a second life.
The blank page came first
Labenz arrived at the microphone with a problem from another business. At Waymark, the video-creation company he founded, the team had made the interface accessible. Customers could use it. Yet some still struggled to make videos because they lacked ideas. As he recalled their feedback: “I still don’t really know what to say.” Making the controls easier had reached a limit. The remaining difficulty lived in the work itself.
Generative AI suggested a different direction: help customers develop the content. Labenz described spending six months concentrating on that opportunity, neglecting other responsibilities until they demanded attention. He eventually handed the CEO role to a longtime colleague and continued deeper into AI. The cost in this account was managerial attention and a change of role. The experience gives his interviewing a practical starting point: where, exactly, does the user get stuck?
The launch newsletter set out an editorial commitment to surface different opinions and follow them as they evolved. Torenberg also named Bankless as an inspiration for connecting immediate developments with bigger debates. Today, the website foregrounds Labenz as its host and “AI Scout.” The name fits the job: go looking, examine the terrain, return with something other people can use.
Follow the pipes
Consider the April 2024 Shortwave interview. Its subject was an email assistant, but the discussion traveled beneath the interface: processing an inbox, choosing a vector database, reformulating queries, retrieving information, reranking results and generating answers. Labenz had used the product himself. He also disclosed that Lee had given him a free year of access. An explanation of the mechanism arrived alongside an explanation of the relationship.
That depth solves a particular information problem. A product announcement tells a reader what the maker wants to sell. A technical conversation can reveal the decisions needed to make the thing work. For a founder or engineer, those decisions may be more useful than the launch itself. For a manager, they can expose the questions an enthusiastic demo has skipped.
“Inspired by what I heard, we started experimenting with a fully agentic approach in Shortwave”Andrew Lee, in a February 2026 testimonial
The archive moves beyond software assistants. Goodfire’s founders discussed mechanistic interpretability, the effort to understand a model’s internal workings. Future House’s Andrew White discussed tools for scientific discovery. A Bolt.new conversation included a browser-based product demonstration. These are different subjects with a shared editorial opportunity: ask the person building the system how its pieces fit together.
The alternatives include Latent Space and No Priors, both familiar options for AI interviews and analysis. The Cognitive Revolution occupies the overlap between a product operator’s curiosity and questions about AI’s broader effects. Its current about page explicitly addresses business, policy and academic leaders. That breadth gives the show room to examine adoption and risk in the same archive.
Free to hear. Work to produce.
The commercial arrangement has two audiences. Listeners receive publicly available episodes; sponsors buy access to their attention. The sponsorship page describes founders, investors, technologists, executives and media operators. Public advertiser listings include Omneky, Brave, Oracle, ElevenLabs, Notion and Box. This is a media business whose distribution crosses audio, video and written updates.
There is an operation behind the conversation. Production partner AI Podcasting describes handling editing, advertising slots, thumbnails, website publishing, articles, release emails and clips. Its case study reports 228 episodes over 24 months, 1.14 million podcast downloads and 10.5 million video views. Those are the partner’s reported results across its engagement, rather than a count of individual people.
Reported by production partner AI Podcasting for its 24-month engagement.
The distinction matters. Downloads, views and subscribers measure different things. They establish distribution; Lee’s account supplies an example of use. Together they suggest why the show can interest both an advertiser seeking a technical audience and a listener trying to make a better decision.
Take a question back to work
The format keeps evolving. October 2026 episodes cover enterprise software reliability and robotics. AI:AM highlights extend the archive with edited live discussions. Their introductions and transitions disclose cloned-voice narration, while recent show notes identify AI-generated summaries. A program about AI is also making AI part of its production process, with the seams labeled.
- Find the constraintWhat made the system difficult?
- Choose one testUse your own task and data.
- Check the resultMeasure quality, time and cost.
For a reader, the useful habit is selective listening. Pick a problem you already have. Use the transcript or chapter markers to find the implementation discussion. Write down one claim you can test, then run a small experiment. That is a way to turn an interview into working knowledge without letting somebody else’s success become your assumption.
The conditions still matter. A workflow built around one model may behave differently with another. Private data, response speed and the consequences of an error can change the answer. An interview gives you access to someone’s reasoning; your own evaluation supplies the missing context. The Cognitive Revolution earns its place when that reasoning is specific enough to travel. The interesting moment comes afterward, when a listener has a better question to ask at work.