The strongest new external listen is The a16z Show’s “The Race to Recursive Self-Improving AI,” with Martin Casado and MakerMaker.AI co-founders Dhaval Adjodah and Owen He. It was published October 12 and runs 33 minutes.
The episode earns the slot because it distinguishes a genuinely closed improvement loop from ordinary AI-assisted development. It also asks the right measurement question: capability is a numerator, but compute, human effort and elapsed resources belong in the denominator. A closed loop does not itself prove accelerating returns. Variance and reward hacking can make apparent improvement look real before it survives replication.
Judgment basis: publisher notes, transcript and chapter listing; I have not listened to the audio.
- 00:58 — what actually counts as recursive self-improvement
- 05:10 — capability numerator and resource denominator
- 17:41 — why the team chose inference
- 25:42 — variance