Published a library doing some very specific data processing. One of the algorithms I implemented was a bit too slow: it would take about a week to process data. I reckon implementation was a little bit sloppy, but I've been implementing a bunch of algorithms from research papers and this was pretty much the published implementation.

I asked Claude to analyse the implementation and check whether it could be improved, half an hour later I got a 26,000% improvement in performance with exactly the same results passing all tests.

Of course, I could have done that myself. But optimization had to go down to simd level; I doubt I would have been able to do that in less than a week of work.