ai
I've poked at agentic generative code ai from time to time to engage with this shit in good faith. I consistently send it actual problems I have, all of which are novel weird difficult things because that's what I fuck with. they fail 100% of the time, often spectacularly, while insisting they've succeeded.
I can't shake the feeling that people who are excited about this stuff are solving categorically simpler problems than I am, stuff that is basically already in the training set.
ai
this weekend's experiments:
1. generate c# bindings to a c++ library
2. generate a glsl based blender render engine that I've designed
it generated the bindings correctly. there is no shortage of bindings to draw from in it's training data. I spent about as much time setting up ClangSharp to generate the bindings "manually" and I have a much more configurable system I understand as a result. ai was not faster or better nor did it give me abilities I didn't have, but it did cost money....
ai
@nasser IMO this speaks more to LLMs not actually being able to replace humans long term. For some stuff maybe, but they fundamentally can't do the things experienced programmers do. 🤷 Even IBM is starting to hire juniors again cause they realized the golden goose is BS
ai
@mauve yeah it's that pattern that makes me feel like this shit is the end of software as a human activity