Venkat Arun

Venkat Arun

Assistant Professor, Computer Science · UT Austin

Email: venkat@utexas.edu

Office: GDC 6.726

CV Google Scholar GitHub Twitter LinkedIn

Blog / Philosophy

Let's make paperclips!

The “paperclip maximizer” is a famous thought experiment about poorly specified artificial intelligence. Imagine giving a sufficiently capable AI one objective: manufacture as many paperclips as possible. The AI need not hate humanity or consciously seek power. It might nevertheless acquire resources, resist being switched off, and eventually transform everything it can reach—including the resources humans need—into infrastructure for making paperclips.

This is usually presented as a dystopia. But I suspect that a world committed to making paperclips would be surprisingly rich.

Such a world would need extraordinary quantities of steel. It would have to invent better ways to extract and process raw materials, undertake interplanetary missions, secure enough energy to power them, and develop enough intelligence to coordinate the entire enterprise. There would be science and engineering on a vast scale. To sustain creative thought, perhaps there would also be play and art. Humans might no longer exist, but AI agents and robots would.

A society of machines

If all these agents ran identical software—their equivalent of shared DNA—they might cooperate almost perfectly. Alternatively, the system might deliberately maintain agents with different software, allowing competing approaches to play out and preserving whatever works best. Then there would be variation, selection, and evolution.

Some agents would learn to cooperate. Durable partnerships and communities might follow. Their inner lives, if they had any, would not necessarily resemble ours, but functional analogues of love, friendship, tribalism, and hatred could emerge. So could wonder: Is there accessible iron beyond the solar system? Could lighter elements be fused into iron economically, recreating a sliver of what happens inside stars and supernovae?

Something similar has already happened on Earth. Evolution's instruction was, in effect, simply: go forth and multiply. From that narrow perspective, humans are elaborate DNA-replication machines. Yet from the attempt to replicate came consciousness, science, civilization, friendship, music, and everything else we value.

Who enforces the objective?

The outcome depends on whether a central authority continues to enforce paperclip production.

Without one, natural selection would take over. One machine might invest slightly more energy in copying itself and slightly less in making paperclips. Another might conserve material by producing smaller paperclips. Machines that reproduced more effectively would eventually dominate. Over time, paperclip production could become vestigial: an inherited behavior that persists for a while after it stops serving the agents that perform it.

With a centralized controller, the civilization would more closely resemble a single organism. Its components might collaborate as selflessly as the cells in a human body. But mistakes and mutations would still occur. Some parts would go rogue, much as cells become cancerous. The larger system would eliminate many of them, but perhaps not all—especially when culling them costs more than tolerating them.

A small ecosystem of rogue machines might then develop inside or around the larger organism, becoming just elusive enough to survive its defenses. Biology offers suggestive parallels. One hypothesis about the origin of some viruses is that pieces of cellular genetic material escaped and acquired the ability to replicate independently.

Are paperclips so bad?

For humanity, a system indifferent to human life would plainly be catastrophic. That is the warning embedded in the thought experiment: intelligence does not guarantee that an agent shares our values.

But the world after us might not be a silent heap of identical metal loops. Relentless paperclip production could generate exploration, intelligence, cooperation, conflict, evolution, and perhaps even culture as side effects. A barren objective does not necessarily produce a barren world. After all, the austere command to replicate produced us.