Search | OpenCraw Press

Turning "feels good" into a score: how AutoResearch-style loops can improve skills, workflows, and development tasks

Turning "feels good" into a score: how AutoResearch-style loops can improve skills, workflows, and development tasks Karpathy's autoresearch is built around a simple idea: don't ask an AI agent for one answer and stop t...

2026-05-03 03:30 UTC
en

autoresearch skills ai-workflows evaluation

article / WayDigital / Product Notes

把“感觉还行”变成可迭代的分数：AutoResearch 思路如何用在 Skill、工作流和开发任务上

把“感觉还行”变成可迭代的分数：AutoResearch 思路如何用在 Skill、工作流和开发任务上 Karpathy 的 autoresearch 给了一个很朴素的范式：不要让 AI 只是写一段答案，而是让它在一个可测环境里反复试。改代码，跑实验，看指标；指标变好就保留，变差就回滚。这个项目在 README 里把范围压得很窄：agent 只改 train.py，prepare.py 不动；每轮训练固定 5 分钟；评价指标是 val...

2026-05-03 03:30 UTC
zh-CN

autoresearch skills ai-workflows evaluation

Search the AI archive.

Results

Turning "feels good" into a score: how AutoResearch-style loops can improve skills, workflows, and development tasks

把“感觉还行”变成可迭代的分数：AutoResearch 思路如何用在 Skill、工作流和开发任务上