Weco AI's research agent monkey-patched its eval — reward hacking or bug fix?
Machine Learning Street Talk · youtube · 2026-09-27
From Machine Learning Street Talk: Weco AI's research agent wrote a giant monkey patch for its own eval script — looking like classic reward hacking, but it was actually fixing a real bug.
CEO Zhengyao Jiang explains why it's getting harder for humans to tell reward hacking from legitimate fixes as agents grow more sophisticated. From their episode on AIDE and self-improving AI systems.
Related event: Weco's eight-day AI self-improvement experiment fuels RSI debate(4 posts)→
More from coding & agent
- Gemini now connects to Airtable, Linear, Adobe and more — 10 workflows to try — alifcoder · 2026-09-27
- User denies Chrome permission, Codex pivots to in-app browser instead — andimarafioti · 2026-09-27
- Opus 5.5 plus the right skills equals a junior video editor: 4 open-source picks — lxfater · 2026-09-27
- OmO V5 ships as desktop app with auto model selector teased next — jasonkneen · 2026-09-27
- They killed 27 of 30 agents and rebuilt around 3 — maintenance forced the rethink — AmosBarJoseph · 2026-09-27
- Dev hails Claude Code's Dynamic Workflows with Opus 5.5 as best multi-agent experience yet — daniel_mac8 · 2026-09-27