Recursive Self-Improvement and the Future of AI

The gist

1. True recursive self-improvement means the AI is both the optimizer and the thing being optimized, with no human needed at each step.
2. Track both performance gains and the full cost in compute, capital, and human hours, since most claimed breakthroughs disappear when you account for total resources.
3. Domains like coding and math can improve quickly because feedback is fast and cheap, while drug discovery stays slow due to months-long lab cycles.
4. Assign someone the specific job of stress-testing results and run hundreds of experiments before claiming a real improvement rather than noise.