Forgive me for being blunt, I can't figure out a more polite way of saying this... since LLMs have trained on enormous amounts of code, and most code is terrible (most of the code I've written is rubbish), why would you expect otherwise? LLMs don't ingest great code, they ingest all code, which on average has the problems described.
Completely agree in principle, I'd expect this when minimizing entropy over any text incl. code. However, evals across variety of domains show that LLMs can reach (and even surpass) expert performance[^1].
