AI for Science: 普通团队进场的位置，不在造神层而在守门人层
Source: https://x.com/snowboat84/status/2078282144619593819
Author: snowboat84
Date: 2026-07-27
Summary (Chinese)
近半年最扎实的 AI4S 入场分析。把赛道按够不够得着切成四层基础设施，用 ScienceAgentBench 约三成、PaperBench 约两成的真实成绩，论证卡住科学 Agent 的不是算力而是学科判断力。给出的筛子：一个位置的胜负由模型强弱决定还是由懂不懂科学决定。结论不是别做，而是做守门人，不做替代者。
Summary (English)
The most grounded AI4S entry analysis in half a year. Splits the track into four infrastructure layers by accessibility. ScienceAgentBench (~30% top accuracy) and PaperBench (~20%) prove that scientific agents are bottlenecked by domain judgment, not compute. Key filter: if a position's outcome depends on model strength, avoid it; if it depends on scientific expertise, pursue it.
One-liner
最实在的机会不在造 AI 科学家，在给 AI 科学家当裁判
*Imported from Obsidian digest notes on 2026-08-01. Content grounded in fetched source metadata via opencli.*