Evidence for RSI: From AlphaGo's Search to Darwin Gödel Machine's 50% SWE-bench
imjustnewatai · x · 2026-08-07
This thread provides concrete evidence for the author's views on Recursive Self-Improvement (RSI):
- Industry Endorsement: Demis Hassabis noted that large models alone are likely insufficient, describing AlphaZero-style planning and search as a "super promising direction." This exact process allowed AlphaGo to discover Move 37, surpassing existing human knowledge.
- Technical Case Study: The Darwin Gödel Machine serves as an early, narrow example. By keeping its foundation model weights fixed while allowing the agent to evolve its own code and workflow, it improved SWE-bench performance from 20% to 50%.
- Caveats: The author clarifies that the 2T vs. 30T comparison and the projected price collapse of intelligence are hypothetical, and notes that search, memory, and verification processes also consume significant compute.
Related event: Beyond Parameters: Recursive Self-Improvement as the Path to ASI(2 posts)→
More from AGI Musings
- Why the Math Problems AI Solves Are Useless to Scientists — stanislavfort · 2026-08-07
- From Economic Agents to Agentic Economies: A Systems Blueprint for Economic World Models — FreedomIntelligence · 2026-08-07
- Data Confirms: China's Tech Optimism is an Outlier Globally — alexmacgregor__ · 2026-08-07
- Quantum Computing Today Feels Like AI Did 5 Years Ago, Inevitable — TansuYegen · 2026-08-07
- Anthropic pays AI chip-design researchers up to $850k, 50% more than its own silicon engineers — teortaxesTex · 2026-08-07
- AI Makes Everything Possible, But Talented People Struggle to Avoid Doing Everything — Dan_Jeffries1 · 2026-08-07