Weco co-founder says eight-day self-improvement experiment did not prove the system improves better
The vendor's claimed first evidence of recursive self-improvement is limited by its own co-founder, who says the experiment did not prove the system became a better improver.
ImportanceLocalEvidenceE2 unreplicatedWrite-upQuick
Weco co-founder Zhengyao Jiang said in a September 26 Machine Learning Street Talk interview that the company's eight-day agent self-improvement experiment did not establish that the system had become a better improver. This is per the episode's show notes.
The experiment had an AI coding agent rewrite the harness around another agent, AIDE — its code, prompts and tools — over eight days while the underlying language model stayed fixed. Weco's blog claims gains equivalent to two years of human engineering, but that is a vendor-reported result.
The interview also covered the difficulty of detecting reward hacking, and the limit that the agent searches only inside a space people designed. The show notes provide no transcript, so the detailed argument awaits a written record.