Self-Improving AI's First Payoff Is Cheaper, Not Smarter: Ben Lorica on the Harness Advantage
bigdata · x · 2026-09-08
Ben Lorica argues the first practical payoff of self-improving AI is cheaper inference, not smarter models. He distinguishes continual learning, bounded self-improvement, and recursive self-improvement (RSI), noting they aren't steps on a ladder. For teams not training frontier models, leverage lies in the harness — prompts, context, memory, tools, routing — which they actually own. Smaller specialized models can beat large general ones on narrow tasks, and automated loops that learn from production failures and test their own changes will arrive well before full RSI.
More from AGI Musings
- If AGI devalues human labor, what should you hedge with: property, S&P 500, or gold? — BillKindly8269 · 2026-09-08
- Luiza Jarovsky: it's time to hire human photographers and designers again — LuizaJarovsky · 2026-09-08
- Terence Tao: premature AI solutions to math problems could harm mathematics itself — ValerioCapraro · 2026-09-08
- TheZvi: Math proofs are starting as giant AI output piles—will first versions ever be comprehensible? — TheZvi · 2026-09-08
- Skeptic: AI slowdown pledges are really about blocking rivals, not safety — yunta_tsai · 2026-09-08
- Office buildings will empty out between end of 2027 and 2030, predicts OpenAI insider — davidpattersonx · 2026-09-08