Ben Lorica: Your Model Is a Rental, the Improvement Loop Is the Asset—Forget RSI Hype
bigdata · x · 2026-09-09
Ben Lorica's Gradient Flow essay separates three conflated ideas in self-improving AI:
- Continual learning: does experience make the system better next time?
- Bounded self-improvement: turning experience into tested, persistent changes
- Full RSI: improving the process that generates improvements
These aren't rungs on a ladder—systems can have one without the others.
Core argument: for most app teams, leverage isn't in the weights (models are rentals) but in the harness—prompts, context, memory, tools, routing, subagents, workflows—which teams actually own. The near-term story is automating pieces of the manual improvement loop: learning from production failures, proposing and testing changes to your own stack. Useful tools will arrive well before full RSI.
Related event: Ben Lorica: self-improving AI's first dividend is cheaper, not smarter(3 posts)→
More from AGI Musings
- Math Bodies Must Decide What Counts as Proof as AI Slop Proofs Flood In — thebasepoint · 2026-09-09
- Ray Dalio: every tech boom creates a bubble — the miracle vs. the investment — RayDalio · 2026-09-09
- AI Video Now Generates Faster Than You Can Watch, Endless Slop Incoming — michalmalewicz · 2026-09-09
- Frontier lab researchers' p(doom) reflects ideology, not expertise, argues gordic_aleksa — D3VAUX · 2026-09-09
- Gary Marcus: frontier labs bet 90% odds of riches against 10% odds everyone dies — GaryMarcus · 2026-09-09
- Why AI lab researchers who believe in ~10% extinction risk keep working there — sebkrier · 2026-09-09