The LLMs are just somewhere between an averaging and a lossy compression of everything on GitHub. There's nothing about the current paradigm of "AI" that is going to somehow do better than just rehashing that training set but with the inclusion of various classes of errors.
I think it's better to view it as spicy search rather than any form of intelligence.
Yeah that's mostly what i use it for. A way to search for things that i can't describe well enough for a traditional search engine so I can find out at a glance what it's name is and if it's valid for my situation. If it is then I go look up documentation. Any time I've stayed in the LLM past that I eventually go down a rabbit hole of wrong ideas that aren't always obvious until you get a bit too deep and you've wasted an hour with an incorrect solution.
Yeah, I do believe it's a good tool for search, just with the caveat that if it can't find an answer it makes one up or otherwise kinda just fills in little missing details with noise.
As I'm slowly evolving my own flavour of spec driven development, I'm starting to think about the generated code as a secondary artefact where main quality criteria is that it's doing what it needs to and it's covered with tests.
I guess my current analogy is that I don't care about how readable or dry is the assembly code generated by compiler.
I have the specifications and the working code with tests. I can always regenerate it if I need to.
But. I still read the produced code, steer the design and correct the obvious blunders. No vibes.
I know a lot of people hate but this AI stuff still isn’t great but it will get better. Each generation of programming languages adds syntax and convenience. AI code will likely get to the point where it is just a higher level language. The only benefit I’m seeing is that if used very carefully I can make more complex projects with fewer team members. And where there was zero documentation there’s at least SOME documentation.
potatopotato
in reply to Arthur Besse • • •The LLMs are just somewhere between an averaging and a lossy compression of everything on GitHub. There's nothing about the current paradigm of "AI" that is going to somehow do better than just rehashing that training set but with the inclusion of various classes of errors.
I think it's better to view it as spicy search rather than any form of intelligence.
Serinus
in reply to potatopotato • • •MrScottyTay
in reply to potatopotato • • •potatopotato
in reply to MrScottyTay • • •Corkyskog
in reply to MrScottyTay • • •doo
in reply to Arthur Besse • • •As I'm slowly evolving my own flavour of spec driven development, I'm starting to think about the generated code as a secondary artefact where main quality criteria is that it's doing what it needs to and it's covered with tests.
I guess my current analogy is that I don't care about how readable or dry is the assembly code generated by compiler.
I have the specifications and the working code with tests. I can always regenerate it if I need to.
But. I still read the produced code, steer the design and correct the obvious blunders. No vibes.
monkeyslikebananas2
in reply to doo • • •utopiah
in reply to doo • • •Might want to read on TDD, it's been around since last the last millennium (OK 1999 according to Wikipedia, point is, it's not new).