Your AI model is a rental, but the loop is an asset: harness-driven self-improvement
bigdata · x · 2026-09-13
Ben Lorica separates three often-confused ideas: continual learning (experience makes the system better next time), bounded self-improvement (the system turns experience into tested, persistent changes), and recursive self-improvement (the system improves the improvement process itself) — not steps on a ladder.
Key points:
- RSI grabs frontier-lab attention, but application teams will see useful forms of automated improvement well before full RSI.
- For most teams the leverage isn't the weights; it's the harness — prompts, context, memory, tools, routing, subagents, workflows — which the team actually owns.
- The near-term story is automating pieces of the improvement loop we currently do by hand: learn from production failures, propose stack changes, test, keep what works.
Takeaway: the model is a rental; the loop around it is the compounding asset.
More from AGI Musings
- Neuralink plus on-demand code models could collapse skill learning to an afternoon — r0ck3t23 · 2026-09-13
- AI feedback in classrooms erases the calibration gap that builds real expertise — DrKavner · 2026-09-13
- Math and AI publishes declaration on AI's impact on human intellectual activity — yoavartzi · 2026-09-13
- Yann LeCun Mocks AI Extinction Scenario as 'Sci-Fi Fan Fiction' — ylecun · 2026-09-13
- Alignment Research Requires Training Researchers, Says Arohan, Pointing to Influence Functions — _arohan_ · 2026-09-13
- Ex-Meta researcher Yacine asks in new blog: why haven't I been replaced yet — yacineMTB · 2026-09-13