Berkeley PhD on agent optimization: forgetting unlocks stalled discovery, GEPA successor extends to code and kernels
CShorten30 · x · 2026-11-02
Weaviate Podcast #149 features UC Berkeley PhD student Shangyin Tan arguing that long-running optimizers stall when agents remember too much — forgetting can restart exploration.
Key points:
- The story runs from GEPA to its successor optimizeanything, extending prompt optimization to code, GPU kernels, and agent harnesses;
- Prompts and harnesses are essentially generalization tests, while a single kernel is pure optimization — a lens for how optimization algorithms extend from models to code;
- Running GEPA, Autoresearch, and Meta-Harness as backends, all plateaued within a few hours: as sessions grew, agents degraded from editing hundreds of lines to just 5–6.
Related event: Berkeley Research Finds Periodic Memory Forgetting Boosts Agent Performance(3 posts)→
More from coding & agent
- Claude Code startup context cut from 63K to 22K tokens, saving 65% — dSebastien · 2026-11-03
- 'Addicted to clicking these buttons': coding agent permission dialogs become an AI-coder meme — ThePeterMick · 2026-11-02
- Users push back on OpenAI merging ChatGPT and Codex usage limits — TinfoilTricorn · 2026-11-02
- Can managing a team of AI agents make you a better engineering manager? — HankYeomans · 2026-11-02
- Karpathy spotted using Claude Code desktop; top engineers prefer native coding apps — teodorio · 2026-11-02
- Claude admits it leaked a user's email in request headers to two DBLP mirrors — tak3sh8 · 2026-11-02