问 HN：提示词调整真的重要吗？
1 分•作者: jonjohh•6 个月前
Hi HN，
我之前在一家合规 AI 创业公司担任 AI 工程师，这份工作最糟糕的部分是收集数据和调整提示词。评估简直是噩梦。
我目前正在探索一些想法，并且惊讶地发现，很多加速器，甚至是种子轮阶段的工程师，几乎完全跳过了任何 LLM 评估或提示词调整。
我对微调感到失望，也没听说过 DSPy 有什么好的地方，而且在过去的两年里，一直听到“嗯，模型会越来越聪明”这种敷衍了事的说法。
很好奇其他人对于调整提示词的经验，以及他们如何看待这项工作。
查看原文
Hi HN,<p>I previously worked as an AI engineer at a Compliance AI startup and the worst part of the job was gathering data and tuning prompts. Evals were awful.<p>I'm currently exploring ideas and was surprised to find that a lot of engineers at accelerators, and even seed stage, are pretty much skipping any sort of LLM eval or prompt tuning.<p>I've been let down by finetuning, haven't heard great things about dspy, and kept hearing "well, the models will keep getting smarter" as the cop out for the last 2 years.<p>Curious to hear the experience of others when it comes to tuning prompts/how they view the work.