LLMs would be 100x easier to build interesting stuff on if you didn't have to optimize for cache hits tbh