RSI Over Scale: How Recursive Self-Improvement Could Collapse ASI Costs

imjustnewatai · x · 2026-08-07

The author argues that achieving Artificial Superintelligence (ASI) may not rely solely on scaling up parameters (e.g., 10T to 30T), but rather through Recursive Self-Improvement (RSI), breaking the link between greater intelligence and massive compute costs.

The core idea is a closed-loop system where a model uses search and planning to test and iterate changes to its own memory, algorithms, architecture, and search processes—keeping verified improvements. A smaller model (e.g., 2T parameters) could eventually outperform a frozen 30T model because its weights are just a starting point for compounding lifetime intelligence.

While total compute demand might explode, the price per useful unit of intelligence would collapse. ASI might arrive from a better cognitive architecture before it emerges from the largest model ever trained.

Related event: Beyond Parameters: Recursive Self-Improvement as the Path to ASI(2 posts)→

Original post →

More from AGI Musings

AGI Musings channel →