> See, AI was trained on existing data - on all that existing C code out there (sure, and also on all the papers and articles saying what was wrong with that C code). Those bugs are in the training data, and often not marked as bugs. So when AI generates C code, is it going to avoid making the mistakes that human code made? No, it's going to generate the kind of code it was trained on. How could it be otherwise?

The generalization of this is why I think all these AI companies writing blog posts where the marketing department is just jer—ranting endlessly about how AI will improve itself into the singularity is just crazy talk. They generate a random statistically likely output, and the most statistically likely output is mid. Exceptional outputs — the ones that wow us or move the needle are exactly that, unlikely. AGI is sci-fi, and LLMs will not change that.

You can see the same effect when AI emits bash, too, and especially so since most bash is terrible, and most users of bash do not put in the effort to learn bash and its foibles. So it outputs what most people write, which is not great.