I don't think LLMs inherently do anything perfectly. They can make sure it compiles and passes tests and they can be trained to do an enormous array of tasks, but the code it generates isn't perfect, it's selecting one of many possible outputs based off of some numbers it came up with after a few matrix multiplications and ReLU activations.

Those matrix multiplications aren't a divine perfect thing. They suffer from floating point precision issues and training data issues and there's still debate if adversarial examples are just an unsolveable property of our linear-algebra based neural network architecture.

Can they do things way faster than a human? No doubt. Can they do very complex tasks? Yes. Do they do things with perfection? Not by our human definition of perfect.