Martin Casado sparks debate over lines of code as a metric for AI recursive self-improvement
wavefnx · x · 2026-11-07
a16z's Martin Casado questioned a large-lab researcher's claim that lines of code (LoC) is a reasonable metric for RSI, citing AI writing lots of AI-building code as evidence. misovalko pushes back with a bell-curve argument: models can produce nonsense in infinite ways but only a handful of right ones; past some hypothetical peak, enough invalid cases could actually make valid ones easier to pin down. What matters is whether the generated LoC carries substantive information. That peak may be task-specific, dynamic, or unreachable — but we still have to climb toward it.
More from AGI Musings
- METR report: Anthropic's AI-driven AI development doubles model improvement speed, alarming researchers — The Verge AI · 2026-11-07
- Box CEO Aaron Levie: long-running AI agents let inference run anywhere, reshaping data center economics — inductionheads · 2026-11-07
- Why is AI treated as inevitable while clean tech gets no such urgency? — _RodolfoOcampo · 2026-11-07
- Yoav Goldberg: AI-generated peer reviews are breaking academic publishing — yoavgo · 2026-11-07
- The Art of Intention: why AI may demand more human creativity, not less — Cyborgized · 2026-11-07
- Musk upgrades to "Universal Extremely High Income," mocked as inevitable branding inflation — JOBhakdi · 2026-11-07