Before the AI craze, I'd gotten quite good at writing regexes. Regexr was quite useful for decoding and composing them. I feel like they're going to become a lost art.

Totally agree. Selfishly, I was always the "regular expression" guy because they were a bit hobby space of mine (engine implementation and such), so seeing LLMs rip them is a bid of a bummer.

Half the reason it's a bummer is because I've seen coworkers who don't know when a regular expression is very suboptimal performance wise, but the LLM has no problem spitting it out. Part of really understanding regular expressions is knowing when to not use them.

The one that sticks in my head is when I was debugging some code that I was suspicious was causing our high memory consumption on a simple API service just to find out the regular expression was being used to strip a potential "data" front of a base64 encoded file (apparently someone thought we should do that instead of rejecting the payload). The regular expression scanned an entire base64 string that was up to 50 MB for the raw file, so about 66MB base64 encoded. I'll tell you what, replacing it with a loop over the first handful of characters solved all the problems. It should've never been a regular expression. If you see regular expressions as an archaic language that solve string problems, and now the magic box can make them for you, you're in for hell.

Regular expressions will always remain fundamental to computer science: They characterise all of those - and only those - conditions on bytestrings (or bitstrings, or Unicode strings, etc) which are checkable in constant memory.* In other words, they characterise the set of all "regular languages", which is a name for DSPACE(O(1)). Furthermore, regular expressions can be matched in O(n) time and O(1) memory, within a single left-to-right pass, which is the highest level of efficiency mathematically possible. Since they operate on bytestrings, they can be applied to computer memory and computer state itself, which are ultimately just bytestrings, and not just to text.

To be fair, you might know all of that, but I wanted to highlight this. LLMs are a lot less efficient than regular expressions wherever both are applicable, simply because everything is less efficient than regular expressions.

* By constant memory, I mean that the memory usage has a maximum value independent of the size or the contents of the input bytestring.