AI編碼智能體可以現代化研究軟體，但無法判斷科學正確性

a field report from openai and academic partners details eight projects where ai coding agents modernized legacy scientific software, ranging from simple build updates to full gpu-native rewrites in rust. results showed dramatic performance gains, with rustqc combining 15 tools to cut runtime from over 15 hours to under 15 minutes, and a gpu-based rewrite running nearly 100 times faster than its cpu predecessor. despite these successes, researchers found that agents like gpt-5.5 and claude code often presented errors with full confidence and lacked the capacity to verify scientific accuracy. this suggests that while ai can handle the heavy lifting of code migration, researchers must still rely on expert guidance and independent testing to ensure scientific validity.