I mean, of course? It's trained on real world code. Real world code "misuses" APIs at, sure, a 62% rate. Sounds roughly right to me. Obviously there's dithe

I mean, of course? It's trained on real world code. Real world code "misuses" APIs at, sure, a 62% rate. Sounds roughly right to me. Obviously there's dithering to be done over what constitutes "misuse" vs. actual bugs, etc... But really this sounds unsurprising. AI isn't going to give you perfect code (or answers, or anything really). It's going to give you typical code, based on extremely broad "intuition" about how others ("all" others, really) have solved (or answered) the same problem. And that has value. But it's not going to produce something better than the existing consensus, by definition.