Discussing RSI: Model Reuse and First-Time Achievement Cost Ratios

willccbb · x · 2026-08-12

The thread explores the evaluation framework for Recursive Self-Improvement (RSI). The author argues that simply comparing "progress multiples" from models is flawed, as much current research inherently depends on good LLMs.

A more consistent metric proposed is the final-run cost ratio between achieving a capability level for the first time versus the second time. For instance, GPT-2 is useless for retraining GPT-2, but advanced models like Fable 5 can be very helpful in retraining themselves.

Related event: Debating the Evaluation Framework for LLM Recursive Self-Improvement(2 posts)→

Original post →

More from AGI Musings

AGI Musings channel →