Feed
The Last AI Built by Humans: Toward Genuine Recursive Self-Improvement
A map of how far AI systems have gotten at improving themselves, in five autonomy levels.
39.9
HCI / 100
Tool-agent headroom closed, 2026
Interactive tool use lags far behind exam-style science.
How it works
Humans run improvement
Humans still pick what to fix, build the data, and check each change, and that cost keeps growing.
Five autonomy levels
Systems are sorted by which improvement decisions AI controls: execution, strategy, experience, deployment, then the improver itself.
Gaps and risks mapped
Most systems sit at L1–L3. Persistent gains need rollback, clear credit for autonomy, and protected evaluators.
From the paper