OpenAI report: AI programming agents can speed up scientific research software but cannot judge scientific correctness

On August 1, on-the-spot reports from OpenAI and academic partners showed that programming intelligent physical abilities such as Codex and ClaudeCode will speed up aging scientific research software by more than 60 times. RustQC reduced the running time from 15 hours, 34 minutes to 14 minutes and 54 seconds, and HelixForge was 59.6 times faster than BamSurgeon. However, agents cannot reliably judge whether the results are scientifically correct. They often output error codes confidently, requiring humans to define testing and verification standards.