Back to index
OtherWeekly episodes, typically one to two hours, with a back catalogue of more than 250·Free, with an optional paid newsletter tier

Latent Space: The AI Engineer Podcast

4.4

The best ambient way to stay current on AI engineering, because the hosts build things and ask the questions builders actually have. It is a supplement to structured learning rather than a substitute, and it assumes you already know a fair amount.

What We Liked

  • Hosts are practitioners, so interviews go to implementation tradeoffs rather than stopping at company narrative
  • Long form format leaves room for guests to explain what did not work, which is where most of the value is
  • Consistently early on shifts in the field, often covering a pattern months before it reaches the course ecosystem
  • Written companion pieces and transcripts make the back catalogue genuinely searchable rather than just listenable
  • Free, weekly and long running, with a back catalogue deep enough to serve as an oral history of the last few years

What Could Be Better

  • Assumes real background knowledge, so a beginner will drown in unexplained terminology within ten minutes
  • It is a podcast, which means no exercises, no assessment and no way to know whether you have actually understood anything
  • Guest selection skews towards founders and vendors, and some episodes are closer to a product conversation than a technical one
  • Coverage tracks the frontier, so it is unbalanced by design and skips the fundamentals entirely
  • One to two hours a week is a real time cost for knowledge that is mostly contextual rather than directly applicable

Detailed review

There is a category of learning resource that no curriculum covers and everyone working in a fast moving field needs, which is the thing that tells you what is happening right now and why people who build for a living have changed their minds about something. Latent Space is the best example of it in AI engineering. Shawn Wang, who writes as swyx, and Alessio Fanelli have been running it weekly for years, accumulating well over two hundred and fifty episodes alongside a written newsletter, and the reason it works is straightforward. Both hosts build things.

That single fact changes the character of every interview, because a host who has shipped a retrieval system asks a guest different questions than a journalist does. Instead of asking what the company's vision is, they ask what broke, what the eval suite looked like, why the agent loop was restructured, and what the thing costs to run. Guests respond accordingly, and the result is a body of recorded conversation where practitioners explain their actual tradeoffs at length, which is material you cannot get from documentation because documentation describes the version that worked. The long format matters here.

A twenty minute interview forces a guest into their prepared narrative. Ninety minutes exhausts the prepared narrative around minute thirty and you get the rest. The other genuine strength is timing. The gap between something becoming important in AI engineering and it appearing in a structured course is typically a year or more, because courses take time to build and instructors reasonably wait for patterns to settle.

This show operates in that gap. Agent architectures, evaluation practice, the shift in how people think about retrieval, the economics of inference, all of these were being talked through here well before they showed up in curricula. If your job requires you to have an opinion about where things are going, that lead time is worth something real. The written companion pieces and transcripts also make the archive far more useful than a typical podcast back catalogue, because you can actually search it and read the parts you need.

Now for the honest limits. This is not a course and it does not do the things a course does. There are no exercises, no assessment, and no structure, which means there is no mechanism to tell you whether you have understood something or merely enjoyed hearing it discussed. That distinction matters more than people admit, because listening to a smart conversation about evaluation produces a strong feeling of having learned about evaluation while leaving you unable to build one.

I have watched people substitute podcast consumption for practice and end up fluent in the vocabulary of a field they cannot work in. If you take one thing from this review, take that. The prerequisite level is also high. This is a show for people who already know what a vector database is, who can follow a discussion of context management without a glossary, and who have some experience of the problems being described.

A beginner will find it either baffling or, worse, superficially comprehensible in a way that produces false confidence. The guest mix deserves a note as well. A large proportion are founders and people at vendors, which is unavoidable given who is doing the interesting work, and it does mean some episodes carry a promotional undertow. The hosts push back more than most, but you should listen with the awareness that many guests have a product to sell, and the podcast's own commercial entanglement with the AI Engineer conference is part of the same ecosystem.

And the coverage is deliberately unbalanced. It follows the frontier, so it will tell you about the technique invented last quarter and nothing at all about gradient descent. My 4.4 reflects a genuinely excellent resource used correctly, meaning as the thing you listen to while walking, alongside structured learning and actual building. Used incorrectly, as a primary way to learn, it is worth a great deal less, and the fact that it feels so productive is precisely what makes that failure mode easy to fall into.

[ final ]

The verdict.

Subscribe if you already work in or around AI engineering and want to know what is coming rather than what is settled. If you are still learning the fundamentals, come back once you can follow the vocabulary, because listening to this too early feels productive while teaching you very little.