@arcana@fedi.layer02.net @lucy@netzsphaere.xyz in my experience llms are mainly useful for: generating comments, generating tests (not the actual contents but a template of basic unit tests that you then write or finish yourself), looking up documentation (just make sure to double check the actual docs afterwards)

@scathach @anemone @lucy @arcana yeah was about to say, that’s a lot of processing power spent on things that can be done better by editor macros

@arcana@fedi.layer02.net @lucy@netzsphaere.xyz yeah generating common patterns often works though it still has a fairly high error rate. I'd compare it to having the average stack overflow poster consolidated into a jr dev employee who works extremely fast but is still pretty dumb

@scathach@stereophonic.space @lucy@netzsphaere.xyz @arcana@fedi.layer02.net usually good enough to describe what a public api does, then I add context if needed

@anemone @arcana @lucy I've used chatgpt to analyze code and make it explain to me what it does, with great success. Not that it was overly compelx code, but still, pretty nice

@scathach @anemone @lucy @arcana they also get it wrong all the time because all they do is pattern matching

@mia@movsw.0x0.st @scathach@stereophonic.space @lucy@netzsphaere.xyz @arcana@layer02.net pattern matching can, imo be useful, but I don't expect LLMS to write code for me, just cut out a little busywork

@arcana@fedi.layer02.net @lucy@netzsphaere.xyz yeah any of the gpt-4 derivatives are significantly better than gpt-3

@paula@netzsphaere.xyz @lucy@netzsphaere.xyz @arcana@fedi.layer02.net yeah im it's not good for writing actual code but it can be good for scaffolding for tests and comments, and so on

@anemone @arcana@fedi.layer02.net @lucy i still don’t expect these to be any better at this task

@arcana@fedi.layer02.net @mia@movsw.0x0.st @scathach@stereophonic.space @lucy@netzsphaere.xyz it depends on what you're going for. You can massage them into writing something that works but it will rarely be done in an efficient or maintainable manner. That's still good enough for some use cases, though, and can save people time if they're not skilled programmers. I worry about the long term consequences of this use case, however.

I mostly fine it useful as a scaffolding (comments, tests) or search/research tool. I can ask it how to do something and the answer will be wrong, but it will likely point me to a useful library, something search engines are pretty bad at these days.

@lucy@netzsphaere.xyz @arcana@fedi.layer02.net yes

@arcana@fedi.layer02.net @lucy@netzsphaere.xyz that wasn't her question though