A technical interview show for people building with AI models. The guest is usually the engineer or researcher who built the thing being discussed, and the questions assume you already know what a KV cache is.
Every episode is pinned to the state of the field on the week it was recorded, and this field moves faster than the archive can age gracefully. The older interviews describe tooling and model generations that have since been replaced. Start with the episodes that explain how a system works, not the ones reacting to a release.
In order. If you only listen to one, make it the first.
Baseten's team walks a 200,000-token request through a production inference stack: cache lookup first, then disaggregated prefill and decode across separate GPU pools, with a speculative decoder trained on coding traffic sitting in front. The best single explanation in the catalog of what happens after you call the API.
Jeff Dean explains how Google runs a frontier model and a cheap one at the same time by distilling the big model down into Flash, using the technique he published in 2014, and why TPU hardware and model architecture get designed against each other rather than in sequence.
Ryan Lopopolo's team at OpenAI shipped a million-line Electron app across 1,500 pull requests without hand-writing code, reviewing it after merge instead of before. The transferable idea: when the agent gets stuck, you go fix the harness rather than write the code yourself.
METR's time-horizon chart plots the length of task a model can complete, measured in human-equivalent hours, and the line has stayed close to straight. Becker is equally clear about which tasks get disqualified from the benchmark, which is where the caveats actually live.
Modal's CTO explains why Kubernetes handles bursty GPU work badly, having been built for web servers that scale slowly, and defends putting hardware requirements and scaling rules in Python decorators instead of separate YAML.
Which is the real problem with "where do I start" lists. Each episode above links to its summary on this site, free and without an account, so you can find out which of the five is worth your evening before committing to any of them.
The bracketed [State of X] recaps taped at the NeurIPS live shows are the most perishable material in the feed, since each one summarizes a single release cycle. The same goes for episodes named after a specific model launch.
The back-catalog problem is the same problem as the queue problem. Listening to summaries instead of full episodes and fitting long shows into a commute cover two other pieces of it.