Almost every AI podcast is an interview show, and almost all of them fail the same way. The host has read a summary, asks a broad opening question, and then reacts to whatever the guest says with agreement and a follow up drawn from a prepared list that does not connect to the answer. The result is a guest delivering the talk they have delivered elsewhere, into a microphone, for ninety minutes. This show does not do that, and the reason is preparation.
The host has read the papers. That single fact changes the shape of every conversation. When a guest describes their approach, the next question is about the specific choice they made in section four and why they made it rather than the obvious alternative. When a guest makes a claim that their own results only partly support, it gets pushed on, politely and specifically.
Researchers respond to this. Someone who has been interviewed several times about the same work, always at the same surface level, visibly changes gear when they realise the person opposite has actually engaged with it, and the conversation becomes the one they would have with a colleague rather than the one they have with a journalist. That is worth a great deal and it is rare. The guest selection is the second strength.
These are researchers, including people working on unfashionable problems, people from linguistics and cognitive science and the older traditions that the current wave has trampled over, and critics who think the field has taken a wrong turn. Compare that with the podcasts that cycle through the same twenty well known figures repeating the same positions. You come away from a good episode here having encountered a way of thinking about the problem that you would not otherwise have met. Because the guests are academics rather than founders, there is no product to sell.
Nobody is subtly steering towards their company's announcement. The conversation can go where the ideas go, and disagreement with the current consensus is allowed to stand rather than being smoothed over. The length is used properly. Two hours on a single research programme means the middle of the conversation can be the technical middle, going into the actual mechanism, the failed attempts, the reason a particular experiment was designed the way it was.
Most long podcasts are long because nothing was edited. This one is long because the subject took that long. Now the problems. The schedule is unreliable, with long stretches where nothing appears.
As a listener you cannot build a habit around it, and it never accumulates the presence in a rotation that a fortnightly show does. That is a shame for the quality of the work and it is the practical reality, and the sensible response is to treat the back catalogue as an archive to search rather than a feed to follow. The episodes are demanding. Ninety minutes to two hours, no editing for pace, and if a tangent goes somewhere unproductive it stays in.
Compared to a show that cuts to forty five tight minutes, this asks a lot, and the reward for the right episode is worth it and the cost of the wrong one is high. The background assumption is significant. There is no explaining from first principles. If a guest works on a particular class of model, the conversation assumes you know roughly what that is.
Someone new to the field will find large parts of many episodes incomprehensible, and this is a show for people who already have the vocabulary. Production is inconsistent. Some episodes sound fine and others carry the audio quality of a call recorded on whatever was available, with the usual distractions of variable levels and connection artefacts. For a two hour listen that matters more than it would for twenty minutes.
My three point eight is for the best prepared interviewing in this format, with a guest list that reaches beyond the usual circuit and conversations that are allowed to be genuinely technical, marked down for an unreliable schedule that stops it becoming a habit, for episode lengths that demand real commitment, and for audio that occasionally makes the commitment harder than it needs to be. Go through the archive, find the episodes on the subjects you work on, and expect to learn more per hour than from anything else in the category.