2026-09-29 · America/Los_Angeles · 社区动态 · #19

Interesting counterpoint to "just scale up": a workshop on optimizing what you already have (DSPy, Oct 3)

Most of what gets discussed here is scaling compute and data. This one's the opposite end of the lever, given a fixed model, how much performance are you leaving on the table because your prompting and evaluation setup is unprincipled. Serj Smorodinsky and Brett Kennedy (co-authors of a book on LLM applications) are running a live session on Oct 3 built around DSPy: defining tasks as signatures/modules, building a real eval set with task-specific metrics, then running few-shot and instruction-level optimization as a formal search rather than manual guessing. MLflow tracks the whole process…

热度 44.8 / 100；排名与评分保留该期记录。