I haven’t seen much discussion about this point other than “llm handle languages x y and z because there’s a lot of training data”. Watching Terence Tau using llm for writing proofs in Lean was a real eye opener in this regard.