The technical consensus that I keep seeing from experienced developers who use LLMs seems to be something like this:
"These things can be useful if you spend a lot of time hammering on them first, already know what you're doing, review their output carefully, and happen to be given access to them or have the money to burn on tokens."
The caveats being:
- People who feel "empowered" by LLMs because they're not developers, are also not likely to level up the same way these folks have because they don't "need" to learn all the things in order to produce something that "works". (For some value of "works" right now.)
- It's very much putting a barrier between haves and have-nots that didn't quite so much exist before. Granted, having (say) fast hardware and such makes development much easier in some respects, but now there's really a barrier of entry in the form of money for Claude or whatever.
- Assumes developers will do the reviews, which is ... optimistic.
It also means that there's much more crap being thrown against the wall in the hopes that it will stick, with maintainers put in the unenviable position of having to wade through it.
I think the long-term effects of this are going to be devastating for FOSS if nothing changes in the short term.
No particularly novel thoughts here, I haven't said anything that others haven't, just organizing some of my thoughts to toss into the void.
All of the "this is just more progress" talk, to me, seems misguided. Sometimes there's a major technological and cultural shift that goes beyond incremental progress and bears worrying about.
ISTM that the jump from automation to LLMs is akin to the jump from conventional weapons to atomic ones.
They represent a massive power shift that has far-reaching, and undesirable, effects. Like the nuclear arms race, the real way to win is not to play, but we just don't seem to be willing to learn that particular lesson as a species.
