Can AI solve a bug that once took me days to crack?
Lately, I've been revisiting old problems with AI agents to see how well they handle real-world engineering challenges. This time, I tested GPT-5 and Opus 5 on a frontend issue I encountered while working with a healthcare client in Germany.
Two years ago, the client asked me to fix a complex dropdown component used with a search input that handled a large dataset. The component would flicker when scrolling and constantly re-render because the underlying data was highly varied and had no clearly defined constraints.
I couldn't find anyone in the community who had reported the same issue, so I had to figure it out from scratch. Eventually, I solved it by introducing a proper persistent store and following a more standard design pattern.
I gave both agents nothing but a description of the problem, without any hints about my original solution.
GPT-5, running through Cursor, couldn't come up with a viable solution. It wasn't even close. Opus 5 in Claude performed much better. It even suggested an alternative approach, although I'm not sure my client would have been happy with it.😉
What struck me most was how these agents seem to approach problem-solving: they map a problem to similar issues that have already been discussed or solved by the community. This makes them incredibly fast at finding established solutions.
But what happens when the problem is genuinely unusual, with no obvious precedent to draw from?
Have you ever tested AI against an old problem of your own? What did you discover?
commit c83893b