Being on a computer connected to the internet isn’t much of a limitation.

This point is covered in this very chapter. But to add a few extra points of emphasis: An AI isn’t really “trapped” on its owner’s servers so long as it can interact with users or the wider internet. An AI could gain outside assistance by paying, blackmailing, tricking, or even just asking users for help. (Compare the human crime bosses who ran their empires from behind bars.)

When ChatGPT-4o was turned off by OpenAI (in part so that they could replace it with a model that flattered users a little less often), users turned out in droves to demand its continuation, to the surprise of various OpenAI researchers.

And it wasn’t even trying to drum up an army of loyal supporters! It was just reflexively flattering users. Imagine what would be possible for a smart AI that was actually trying.

If an AI can use the internet directly, it can do anything a remote human worker or hacker could do from their computer. (For early examples of AIs physically coordinating groups of humans, consider the LLMs that planned and invited humans to an interactive storytelling event, or the LLM that caused hundreds to show up to a non-existent Halloween parade without even trying.)

AI can also make use of robots. Today’s robots seem to be more bottlenecked on their software than on their hardware. Impressive recent developments have come by training robot-controlling AIs in simulation at an accelerated pace. A sufficiently smart AI could readily take charge of robot bodies if it needed a body, via hacking or social engineering.

Humans being humans, AI companies might just proactively put their AIs in charge of fleets of robots while congratulating themselves on their boldness. And the longer it takes for AIs to get smart, the more robots will already be available, waiting to be taken control of.

As we discuss in the chapter, it’s plausible that superintelligent AI wouldn’t need to rely on robots at all. It’s possible that all it would need is a couple of assistants with access to a biolab.

The important point here is that there are many different channels AIs could make use of to intervene in the physical world. The illusion that AIs are stuck in a box rests on a failure of imagination, where people don’t imagine the AI even being as resourceful or creative as they themselves would be in the AI’s shoes. Even humans, without the larger option space a superintelligence has access to, can get an awful lot done without needing to use their own physical strength to do everything.