Counter point though- what if it was trained on your specific code base? Wouldn't it be able to then help with those given nuances?
The code base I have I would love to be able to just give some AI free reign and learn the structure since a lot of it is fairly repetitive; I know it would be so easy to say "hey add X just like Y" and it would be able to do it easily.
The rare times where I have no good idea what to do, it is faster to code with an LLM. The rest of the time, when I know just what I want, it takes longer for me to formulate it, query the LLM and discuss and validate its output.
My more experienced/senior colleagues all say roughly the same. It’s great help for our juniors though. They learn a lot and are more capable on their own with the AI assistance.
It’s improving all the time though, so I’m not writing it off at all. I am developing an evaluation suite so I can keep watching the progress in a systematic way..
Open benchmarks are vulnerable to saturation. I think benchmarks should have an embargo periodic, until which only 3% of the question-answer pairs is released, with an explicit warning not to use it 3 months after being released.
I think there are types of problems that AI will be great at solving. If you can pass in a whole codebase then we might have LLMs that can suggest refactoring etc to improve code quality.
Code mods like upgrading from python 2 -> 3 could also become possible.
The code base I have I would love to be able to just give some AI free reign and learn the structure since a lot of it is fairly repetitive; I know it would be so easy to say "hey add X just like Y" and it would be able to do it easily.