Weco AI Workshop Breaks Down AutoResearch Loop Behind OpenAI's Parameter Golf Challenge
AI Engineer · youtube · 2026-10-12
At the AI Engineer World's Fair, the Weco AI team (Vayum Arora, Zhengyao Jiang, Dixing Xu, Dhruv Srikanth) ran a hands-on workshop on Parameter Golf: an agent loops over code-change proposals, keeping edits that improve a model's score and discarding the rest.
Key points
- Zhengyao Jiang frames autoresearch as repeated proposal-and-evaluation loops
- Dixing Xu covers the Aiden system and its contributions to OpenAI's Parameter Golf challenge (16 MB total weights, 10-min training budget, BPB metric)
- A practical recipe: one modifiable source file + one fixed evaluation command + a measurable goal
What makes a search useful
- Fast, stable evaluations; a clear metric direction; time limits; initial guidance; protection against reward hacking
Live demo
- Weco's experiment tree makes each attempt inspectable (code diff + score in the dashboard); you can branch from promising work or steer a running search toward a new idea
- Mini Parameter Golf is open source (github.com/WecoAI/mini-pg), experiments run on Modal
- Final exercise swaps model training for a Flappy Bird controller scored across normal play, moving gaps, wind and mixed conditions, with a hidden validation set — a lesson in designing evaluations that reward the behavior you want
Q&A touched on reinforcement learning and combining multiple objectives.
More from coding & agent
- Handy Claude Code trick: just reference @AGENTS.md from your CLAUDE.md — weswinder · 2026-10-12
- Custom church payment system: GPT security audit caught a critical flaw before it was too late — petergyang · 2026-10-12
- AI Agent Does CAD: KiCad MCP Reads Mounting Holes for a Robot Build — burhop · 2026-10-12
- 324 chess games show LLM self-notes act as memory but can't fix board blindness — marvijo-software · 2026-10-12
- Opus 5.5 Compared: Claude Teammates Push Back, Codex Quietly Drifts Behind Green Tests — Sauers_ · 2026-10-12
- AI Writes Code Faster — So Why Aren't We Shipping Faster? — kristiyanstoyanovAI · 2026-10-12