With computers, the hard part is getting them to solve a certain problem at all. High volume and speed come soon after.

“To take over the world, you need a civilization” is an intuition that makes sense for humans. It’s a lot less obvious how well this idea generalizes to AI. AIs don’t work like humans — they can be dramatically more capable than any human, and an AI instance isn’t necessarily comparable to a single person.

It’s also worth keeping in mind that superintelligence is exactly the kind of thing that can end up with an analog of a whole civilization extremely quickly.

With most feats that computers can manage, it doesn’t take long to go from “computers can do this” to “computers can do this at an enormous scale, far faster than any human can.” Think, for example, of calculators.

There were years when only the highest-end computers could do speech recognition, video processing, or real-time 3D graphics, but there weren’t very many such years.

AIs, like traditional software, can be swiftly copied onto as many computers as are available. And more computers can be built at the speed of industry.

Compare this situation to humans. Creating and training up a new human takes substantial resources and decades of time. Once you have a single AI at a given capability level, you can immediately copy that same trained, “adult” AI as many times as you want, at minimal expense.

In a sense, a whole (small) civilization’s worth of AI minds already exists the moment a company rolls a new model out to their datacenters and spins up as many instances as needed to fill demand. Today, those AI fleets aren’t all working in harmony. But companies do use groups of parallel agents when aiming for the highest performance at any price.

This all means that there likely won’t be all that much time between when AIs become smart enough that they could take over if they had a million instances and when AIs have at least that many instances running. The kind of population growth that takes humans hundreds of years can occur in minutes with AI.

When it comes to the physical infrastructure of civilization, we would guess that AI can productively piggyback on human infrastructure for however long it takes to develop a more advanced means of rearranging matter to its preferences. It doesn’t need to figure out how to manufacture its own supply chain and computing infrastructure from scratch when it can use our computers. It doesn’t need to invent industrial machines from scratch when it can just take control of industrial machines that we’ve already helpfully built. And it can use our infrastructure to build the next phase of its own infrastructure, using existing robots to build new and more efficient robot factories, or using existing DNA synthesis laboratories to make its own biotech, until it’s completely self-sufficient.

A handful of humans started out naked on the savannah, and we bootstrapped our way up to a technological civilization. And we’re not that smart. It wouldn’t be especially hard for a superintelligence to do its own form of bootstrapping, particularly if it gets to start from humanity’s existing industrial base as a leaping-off point.

Notes