I can't even imagine what "next token prediction" would look like generating x86 asm. Feels like 300 buffer overflows wearing a trench-coat, honestly.