The fact that we're willing to rely on LLMs to generate code knowing that it only mostly works because that code has been generated before thousands of times is not an indication that stochastic models are good, it's a massive, punishing indictment of computing as a field.
Using an insanely huge expensive model to quasi-reproduce work that's been created thousands of times already isn't productive or efficient. It's a symptom of profound failures of language, practice, process and imagination.
The one bright spot of LLMs is that we've learned exactly how morally and creatively bankrupt this industry is, and how fast people are willing to throw away everything they said they cared about for decades - craft, understanding, efficiency, detail, all of it - just to cosplay competence and bandwagon their way to groupthink targets they don't even realize were made up to manipulate them.
Christ it's embarassing.