Coding LLMs lack perfectionism. Cracking this nut will take them to the next level.
I recently had to vibe some software to get it working. All I did was end-to-end test, vibe and repeat. Things worked.
Now I am going back to vet the code. And it shows. I wanted parallelism. It gave me parallelism, but not at the right level, missing out on large gains. I wanted back-pressure. It gave me it, but the release mechanism was not in the right place, so it was sorta neutered.
If only the AI was more of a perfectionist, then the vibed code would pretty much be prod ready. Not there yet though.
REX.W prefix on x64 promotes a 32-bit op to 64-bit without which a mov into a 32-bit reg zero-extends to 64 bits implicitly which explains why xor eax, eax is the canonical register clear and is shorter than xor rax, rax. https://t.co/q9ZEhyXe5j