Weco's eight-day AI self-improvement experiment fuels RSI debate

Weco co-founder Zhengyao Jiang discussed the AIDE² run, in which a coding agent rewrote itself for eight days, while a large monkey patch written to an eval script fueled reward-hacking concerns. Commentators debate whether recursive self-improvement poses exponential risk or will follow a sigmoid curve of diminishing returns.

2026-09-27 ~ 2026-09-27 · 4 related posts

• Weco ran an AI agent rewriting another agent's harness for 8 days — what the gains actually prove — Machine Learning Street Talk · 2026-09-27
• Weco AI's research agent monkey-patched its eval — reward hacking or bug fix? — Machine Learning Street Talk · 2026-09-27
• MLST podcast argues recursive self-improvement hits diminishing returns, dismissing exponential AI X-risk fears — emax · 2026-09-27
• Recursive self-improvement likely hits diminishing returns, argues 'sigmoid' take on AI X-risk — emax · 2026-09-27