claude is kinda smart
i asked the computer to do something and it did. no notes (well, some notes below). this makes me think
the problem it solved
i like to play n++ from time to time. it's a challenging platformer game which is notable, among other things, for it's choice of music. unlike other entries in the genre, which come with music created specifically for the game, n++ features a playlist of existing electronica tracks.
sadly, one of the tracks, which was also one of my favorites, has been removed due to licensing issues (apparently it's too expensive to keep the license going) so an update has been released which removes this song from the game. this sucks, so i put it back.
my first attempt was to obtain an older copy of the sound bank which the game uses to store music and drop it in, but that didn't work. clearly, the game hard-codes somewhere which tracks from the sound bank are actually played.
this is where claude comes in. i opened claude code in the installation folder and told it to get to work. every time it patched the game, i gave short written feedback, nothing more than e. g. “you're done”, or “it's not showing up”.
after three rounds of this, claude had figured out the memory layout of the song list in the game binary, and written a patch that puts the song into an unused slot of that list and makes that slot used everywhere.
learning the skills required to do this would've taken me longer. i am vaguely aware of reverse engineering but actually coming up with a patch for a complex binary has been beyond me so far. this is kind of crazy and meaningful imo.
so what, will it replace everyone? will it ✨agentically transform software engineering✨
eeeehhh, maybe. ai is clearly sometimes good at “software” if that can be broken down into one-shot tasks, as above. in the case above, if i wanted to do more (like add another song that was removed back before i even got the game) it was already getting measurably more expensive to continue the session and i don't think i'd be willing to pay extra by a lot to do so (or orient my sleep around waking up when the rate limit resets)
also, there is a hard limitation on “context”, that can be fed into these ai models, so current models leave their users no choice but to do some sort of reset from time to time. there are mechanisms that attempt to overcome this, but it's not like the ai can currently “learn” from completing a task like this in the same way a human can.
this does not seem like an unsolveable problem though, and alongside people putting up a clear display of insanity about it, there are some people on social media who i kinda trust on this who are working out similar approaches
it certainly makes me feel things about my life choices, and i don't think it can be readily dismissed as just some toy. obviously some of the ways people use it are irrational, but that's true about most things.
on language
i've seen some people argue that it's wrong to say that ai software is “thinking” or that i “asked” it to do something because it supposedly humanizes ai, which they think is immoral to ever even do in words. i don't think this is a reasonable complaint, even if you vehemently disagree that llms resemble human thinking to the point where you find language which makes that assumption objectionable.
imagine i'm using a website where there is also a mobile app. now assume there is a bug where the website shows information a' but the app shows the outdated information a. “the app still thinks a is true” would be a reasonable (if slightly informal) way to describe this bug, without anyone ever assuming that this software system is conscious or human or whatever.