AI Can Write Supercomputer Code — But Still Needs Human Help

Supercomputers power things you use daily: weather forecasts, airplane design, and drug research. But writing software for them is incredibly hard and requires rare expertise. A new survey tested whether AI language models (the same tech behind ChatGPT) could help. The answer: they can be useful assistants, but they’re not ready to work alone.

The study grouped AI’s supercomputing skills into five areas, including writing code, speeding it up, and checking quality. The good news is AI handles simple serial code surprisingly well — think of it as writing a recipe from a list of instructions. AI also does okay with shared-memory tasks like OpenMP, where multiple cores work on the same data. But when supercomputers split work across thousands of processors (a technique called MPI), AI stumbles. Errors here can crash a $100-million machine or produce wrong scientific results.

Specialized AI models trained specifically on supercomputer code — with names like HPC-Coder and HPC-GPT — performed better. They use clever tricks like looking up documentation while generating code, similar to a chef referencing a cookbook. But these tools are narrowly focused and mainly tested on small sample problems, not real-world supercomputing projects. In production, small mistakes become huge, and performance can't be compromised.

The researchers conclude AI won't replace supercomputing experts soon. Instead, the future is collaborative: AI handles routine coding, saving scientists time, while humans double-check correctness and optimize performance. For you, this means faster discoveries in medicine and climate science — but it will take years of careful work to make AI reliable enough.