Craig Smith spent years as a New York Times correspondent and brings the habits of that job to this podcast, which mostly works in its favour. He reads the papers, he knows who the guest is and why they matter, and he asks questions that show preparation rather than a producer's list. In a field where most interviewers have read the press release, that stands out. The guest list is the reason to listen.
Smith has had on researchers whose work genuinely shaped the field and who almost never appear on podcasts. Yann LeCun, Gary Marcus, Jürgen Schmidhuber and a long list of academics and lab researchers who are influential without being famous. Some of these conversations exist nowhere else. If you want to hear a particular researcher explain their own position at length, there is a reasonable chance this show has done it.
The format gives people room. An hour with one person and no segments means a guest can develop an argument properly, including the parts that require setup. Contrast that with a fifteen minute news podcast slot where everything gets compressed into a soundbite. Ideas that need time to explain get time here.
The show largely ignores the news cycle, which I appreciate more than I expected to. There is no scramble to react to the model that shipped on Tuesday. Conversations are about research directions, about disagreements within the field, about what a particular line of work is actually trying to achieve. That gives episodes a longer shelf life than most.
The coverage of non mainstream directions is valuable. World models, neurosymbolic approaches, alternatives to the scaling hypothesis, robotics, neuroscience adjacent work. If your entire information diet is language model discourse you will have a distorted picture of the field, and this show corrects for that better than most. Now the problems.
The audio is genuinely inconsistent. Some episodes sound fine and some sound like a bad video call recorded through a laptop microphone. For a show whose whole value is listening to people talk, this matters, and it is the single thing I would most want fixed. The interviewing is generous to a fault.
Smith gives guests space, which is a strength, and he rarely challenges them, which is not. When a researcher makes a strong claim about what their approach will achieve, or dismisses a competing line of work, the follow up is usually an invitation to say more rather than a request for evidence. The best interviews in this genre have some friction and these mostly do not. Pacing is slow.
Episodes meander, there are digressions that go nowhere, and a forty minute conversation sometimes contains fifteen minutes of substance. Higher playback speed helps. Sponsorship is present and it appears to influence some guest selection. There are episodes with enterprise AI vendors that feel materially different from the research conversations, and they are the weakest thing on the feed.
They are easy enough to identify and skip. There is no on ramp. No introductory episodes, no explanation of terms, no sequence. You need enough background to follow a technical conversation between two people who both already have it.
A newcomer will not get much from this. My three point nine is for genuinely valuable long form access to serious researchers, prepared interviewing, and a refreshing indifference to hype, marked down for audio that undercuts the whole proposition, an interviewing style that does not press hard enough, and slack pacing. Browse the back catalogue by guest name. There is a lot in there that is worth your time.