Software speed and responsiveness are broadly underrated.
Many users would like it if their software were faster, but they won't say "I'd like it if this were faster." This is partly because many users don't explicitly understand all the ways in which software can be fast and responsive, and partly because "it should be faster" is just not a common thing less-technical people say when you ask them about software.
AI also helps here because many performance problems can be addressed by using slightly less bad techniques, and AI equips you with a partner that knows lots and lots of reasonable techniques.
The software I love (e.g., Drafts) is, disporportionately, highly performant.
I broadly decompose performance gains into (i) basic improvements that are often simplifications and rarely have significant tradeoffs, and (ii) really complicated optimizations that add significant complexity. The "Complexity" section of this Dan Luu study is a good discussion of (ii).
Because category (ii) exists, and because performance problems require a lot of judgment in general, simply asking AI to "make this faster" can cause big problems (at least as of June 2026). A longer conversation where you go through the profiling and consider options tends to work much better.
Because category (i) above exists, studying your software's performance "prematurely" is a good way to discover basic structural problems. Yes, "premature optimization is the root of all evil," but what looks like optimization is sometimes just a way to audit your software for horrible problems.
I've sometimes felt intimidated about working on performance, because I associate "performance work" with game development, low-level programming, and other subfields full of people who know orders of magnitude more than I do about cool and difficult subjects. Remember that a lot of real-world performance work is not optimizing ray tracing but instead "hey, we were blocking on X for no reason and now reads are six times as fast."
But even if you are not anywhere near game development, reading about (e.g.) ray-tracing hacks is valuable for building intuition about computer systems in general. It's also great fun. I know very little about allocation, but still found this piece making the rounds very useful.