利用 Codex 加速科学研究
Accelerate scientific research with Codex
已完成 Codex 和 ChatGPT 的工作。一位研究人员广泛使用 Codex 和 ChatGPT 进行科学研究，一天内完成了一周的工作，尽管最近的用法限制减少了输出.
做了什么
A theoretical chemist says they used ChatGPT and Codex heavily in a research workflow built around quantum-mechanical simulations of molecules, excitations, and spin behavior. In a follow-up comment, they say they reworked the setup to use one model for most analysis, another as a final reviewer and coordinator, and a cheaper model for routine jobs like starting and stopping calculations, copying files, and parsing data, while they chose when to escalate harder reasoning.
效果如何
They report unprecedented productivity over the past year, saying the tools let them do far more research than in 10+ years before, but that recently the same work now burns through a weekly allowance in a day. OpenAI later told them a bug affected their usage and gave them $500 in credits; they say they had previously burned through $3,000 in credits quickly too.
值得注意
The main catch is cost and limits: they now check cost-to-performance before each run and say they have had to scale back usage because the expense is no longer sustainable.
试试这个 Prompt
I'm researching [scientific question or topic]. Using the data and code in [folder path], reproduce [analysis or result], then run [next experiment or analysis] and summarize the findings with plots saved to [output folder]. Keep a log of assumptions and anything you couldn't verify. You're done when the scripts run end to end and I have a short write-up of the results and open questions.