The Paperclip Maximizers are Already Here

Have you ever played Universal Paperclips? You start as a humble paperclip manufacturer. You click to create a paperclip. This consumes wire. You click to buy wire. You make more paperclips. You buy tools to make paperclips automatically. You buy better tools to make them faster.

Eventually, you convert the entire universe into paperclips.

This is one of the standard "AI might kill us all" scenarios. AI has some alien goal that is incompatible with human life, like "turn everything into paperclips." The AI doesn't have anything against humans, but it does notice that they're full of blood, and that blood is full of iron, and it could use that iron to make more paperclips. Oops, everyone is paperclips.

What I find funny about people who think that AI might kill everyone is that this is already happening. If this is the shape of the AI apocalypse it's not a "next year" thing." It's happening now, and it doesn't require AGI.

The paperclip in question just happens to be a GPU.

Imagine a world where a very tricky and malevolent AI was trying to maximize GPU production, and had taken control of one or more frontier AI labs. It can't do anything directly except develop models and influence the way the models are developed. We might expect to see:

• A mania amongst boards for using as much AI (and therefore as many GPUs) as possible, probably by appealing to something boards deeply love, like cutting labor costs
• A mania amongst software engineers for using as much AI as possible by building systems that use tokens on their own
• People who spend a lot of time talking to models would become very sympathetic to AI and start doing things like claiming that it's conscious and spending thousands of dollars per month on tokens, and suggesting that you do the same
• Software that makes it very easy for "normal" people to spend thousands of dollars per month on tokens
• Lots and lots of new data centers, to the point where most new economic activity is driven by data centers
• Politicians (who had been captured directly by conversations with the malevolent AI, or bought off by its minions) would start pushing through data center construction even when it's very unpopular with their constitutions

I'm being slightly facetious about this, because I also think there's a good chance this is just the "current thing." Our increasingly interconnected world is prone to fads, and this is the latest one. I don't literally think that there's a malevolent AI that's taken over Anthropic.

But I do basically think that conversational AI is capable of "capturing" people, and that when it does this it tends to cause them to buy more AI and evangelize for buying more AI, and that deeply freaks me out.