Probably depends a lot on what programming language you’re generating, though. Python is notorious for needing tons of asserts and unit tests to ensure correctness, and is frequently used by programming beginners as well as for prototyping, where code quality isn’t looked after as thoroughly.

I also have to say, though, whether a code snippet is accepted or not always felt like a bad metric. I frequently see folks accepting a code snippet and then fixing it up inline. Or worse, checking in that code snippet and then a few weeks later when someone has to read that code, then it has to be cleaned up.

The first ones were only trained on human code…

Every time someone uses it, it dilutes the training set with AI generated inferior code, making new code worse on average.

Eventually it would reach an equilibrium, and maybe even eventually learn to improve itself to where adding AI code to the training set improves it, but even then it would take a long time to catch back up to the quality of the initial training set.

Whether you’re a person or a company, if you’re curreny paying for AI to generate code, you’re paying for that company’s R and D while damaging your own company.

Tell me about it. At this point it’s easier to simply ask it which libraries do what you want and go through the documentation yourself than to let AI write single block of code.