You’re being intentionally obtuse here. Right now LLMs are harmless because we only let them print characters to the screen. But suppose you have an assistant version that you allow the ability to execute code. And you ask it to write some code to process an excel file and run it, but while it does that it also decides to copy itself to an external server you don’t know about and starts doing anything there. Without reviewing every thing it does, you can’t be certain that it’s not doing something malicious. But if you have to review everything it does, then it’s not nearly as powerful and helpful for automating tasks as it could be.

You say you can destroy it by destroying the computer it’s on. But you can’t do that. You have no idea what or where any given EC2 instance is located, and if you did, you wouldn’t be able to get there before the AI transfers itself to another computer within a few minutes or seconds.

A truly rogue, intelligent, sentient AI hell bent on damaging the world, unleashed onto the internet could do untold damage to our society.