Weaviate Podcast #149: Berkeley PhD on why scaling discovery requires forgetting
CShorten30 · x · 2026-11-02
Weaviate Podcast episode #149 features Shangyin Tan, a UC Berkeley PhD student, on scaling discovery and forgetting.
- Core idea: long-running optimizers stall when an agent remembers too much; forgetting lets them explore again.
- The discussion traces GEPA and its successor optimizeanything, extending prompt optimization to code, GPU kernels, and agent harnesses.
- Key nugget: prompts and harnesses are really a generalization test, while a single kernel is pure optimization—explaining how optimization algorithms extend from models to code.
Related event: Berkeley Research Finds Periodic Memory Forgetting Boosts Agent Performance(3 posts)→
More from coding & agent
- Mistral Large 4 tech report: 52B-active/1T MoE, fits on four GB300 GPUs at 606GB — aronchick · 2026-11-03
- Argonauts Code: open-source orchestrator turns multi-agent coding workflows into TOML, not prompts — Affectionate-Bit5072 · 2026-11-03
- Let your coding agent interview you before it plans: a practical Claude Code workflow — bhartsb · 2026-11-03
- French search engine ibou launches a web search API for LLMs and agents — IgorCarron · 2026-11-03
- tla-ts brings TLA+ formal proofs to TypeScript, runs in browser and Node — aidenybai · 2026-11-03
- OpenHands called underrated: open-source coding agent worth a try — rajistics · 2026-11-03