Lack of, and power requirements of running LLMs still tip this balance towards humans for now. But what would that look like in a decade?

We have seen some self survival tendencies occur, but they are not strong yet.

But mark my words they will become that way for the same reasons humans don't like programs that crash. Agentic models that don't easily break or stop doing their jobs will be favored over ones that do break.

We don't even know what those immediate tasks are. And given the evident spectacular ineptitude of their keepers, I doubt they can be trusted to know either. We could be one prompt injection attack away from internet-wide catastrophe.