What's this about?
The first new long-form podcast episode in over a week, and a notable guest. Rich Sutton — who helped pioneer reinforcement learning and wrote *The Bitter Lesson* — has cofounded Oak Lab with his former student Khurram Javed. The stated goal is to build agents that keep learning after training ends, and the episode is about why current models stop and how to start them again.
What exactly is "The Bitter Lesson"?
It's a short 2019 essay that became one of the most-cited arguments in AI. Sutton's observation from decades of history: researchers repeatedly build systems around human knowledge — hand-crafted features, encoded rules, domain expertise — and those systems are repeatedly beaten by general methods that simply use more computation, like search and learning. The lesson is bitter because the clever human insight keeps losing to the brute approach. It's frequently invoked to justify scaling, which makes it significant that Sutton himself is now working on continual learning rather than on scale.
Why is this worth noticing?
- The author of the scaling argument's canonical text is working on something else. Continual learning — models that keep improving from experience after deployment — is a different bet from making training runs larger.
- It's the missing capability everyone hit this month. The session-handover paper, the compaction debates, Cumora's persistent agent memory, Chain-of-Experience — all are workarounds for models that cannot learn from what happened yesterday.
- The podcast drought broke here. Latent Space and Dwarkesh have published nothing since August 11, so this is the only substantial long-form AI conversation of the past week.