I got a Claude Max subscription and have been playing around.

First big experiment was to see if we could reimplement a part of our pipeline from Julia into Rust. The quality of the code it produced was really good, and it did good analyses on what parts would be easier or harder to port. Eventually this convinced us that a port wouldn't have been worthwhile, because the Rust <> Julia interop is currently quite difficult. I think it might have taken us several days to come to that conclusion otherwise, so this was a big win.

Second significant experiment was asking Claude Code to optimize a pipeline. It pretty much failed abysmally – it successfully added a lot of profiling and came up with a lot of ideas, but not a single one of them made the code any faster. Later, I was able to make a significant improvement to that pipeline with about half an hour of work. So, a failure here.

Third was a moderately difficult refactor, replacing some dicts with structs, that then get serialized and read in another part of the codebase. Claude Code did really well here – it occasionally produced code that was overly verbose, but nothing was *wrong*, and it took corrections/feedback quite effectively. Another big win here.