Hot take: narrow recursive self-improvement has already begun inside LLM pipelines

maksym_andr · x · 2026-09-23

A contrarian take arguing that narrow recursive self-improvement (RSI) is already underway: LLMs routinely optimize their own inference, generate new RL environments, curate pretraining data, write experiment code, and launch training runs. Full end-to-end RSI is technically feasible now, but nobody seriously tries it because the improvement slope without humans in the loop is much worse — plus safety concerns. The real question, the author argues, isn't whether RSI happens but the rate of AI-alone progress versus AI+humans.

Original post →

More from AGI Musings

AGI Musings channel →