That's been my main argument for why LLMs might be at their zenith. But I recently started wondering whether all those codebases we expose to them are maybe good enough training data for the next generation. It's not high quality like accepted stackoverflow answers but it's working software for the most part.