Superintelligence: The Footprint Problem

Listen to Free Audio Version Here

By Charles Richard Walker

A pause is a strategy for the present industrial shape of artificial intelligence. It is not necessarily a strategy for intelligence as a category. Right now, the frontier of AI looks something like a factory. The most ambitious systems require enormous computing clusters, advanced chips, substantial electrical power, specialized personnel, sophisticated cooling and networking infrastructure, and training operations concentrated in facilities that governments can identify. That physical reality is what makes proposals such as chip tracking, export controls, datacenter monitoring, licensing, international inspections, and restrictions on enormous training runs plausible in the first place. Governments cannot see an algorithm thinking, but they can see the industrial infrastructure required to run it.

The mistake would be assuming that this industrial shape is permanent. Cost is not a side issue in the artificial-superintelligence debate. Cost is part of the control surface. Almost every proposal for preventing the construction of an extremely powerful AI depends somewhere upon the assumption that the dangerous step remains expensive, centralized, scarce, and conspicuous. When that assumption eventually fails, because all technology gets cheaper over time, many of the proposed controls fail with it, even if every government enforcing them acts in good faith.

The important question is therefore not simply how much computation today’s frontier model requires. It is how the minimum resource bundle required to perform increasingly capable automated research changes over time. Hardware improves. Algorithms improve. Training methods improve. Models can be compressed, specialized, fine-tuned, distilled, and supplemented with tools and inference-time computation. Knowledge spreads. Techniques that initially require teams of specialists eventually become standardized procedures. The original achievement can be enormously expensive while reproducing portions of its useful capability becomes progressively cheaper.

The argument is not that tomorrow somebody will train the equivalent of a frontier model in a basement by a kid with Cheeto crumbs on his T-shirt, although it might. It is that the amount of money, hardware, energy, expertise, and physical infrastructure required to produce a given amount of useful machine intelligence can decline. If that process continues far enough, the threshold eventually moves from something only states and frontier laboratories can attempt toward something accessible to universities, ordinary companies, smaller organizations, and perhaps eventually individuals or distributed networks. At that point, prohibition becomes a fundamentally different problem, if not impossible.

This becomes especially important when the thing being made cheaper is research itself. The critical threshold need not be the sudden appearance of a godlike machine capable of redesigning itself overnight. Recursive improvement could begin in a much more mundane form. AI systems become better at programming. Better programming systems assist with AI research. Research systems help design experiments, search architectures, diagnose failures, optimize inference, generate synthetic data, improve training procedures, and perhaps eventually contribute to hardware design. Better tools then assist the humans and machines developing the next generation. The loop does not initially need to exclude humans. It merely needs to exist: better models produce better research tools; better research tools make the next improvement cheaper or faster; the improved system becomes an input into producing the system after that.

The first turns of such a loop could occur inside enormous laboratories. There is no guarantee that later turns would require laboratories of the same size. This is where the physical footprint becomes crucial. Much of what makes frontier AI governable today is precisely what makes it inefficient enough to be visible. If producing a certain level of capability requires tens of thousands of specialized processors and extraordinary amounts of electricity, the activity creates something governments can potentially detect and regulate.

If future algorithms achieve comparable research productivity with one-hundredth of those resources, the footprint contracts. If another architecture eventually achieves it with one-ten-thousandth, the problem changes again. At some point the relevant comparison may cease to be enormous datacenters and begin to involve university clusters, commercial servers, distributed networks, compromised machines, or hardware categories that governments never designed their original controls around.

This is why controlling chips is fundamentally different from controlling knowledge. A government can restrict a particular accelerator. It can monitor a fabrication plant. It can regulate a datacenter. What becomes extraordinarily difficult is preventing civilization from retaining the knowledge that a more efficient method exists once that method has been discovered and disseminated. Knowledge does not conveniently un-invent itself. A moratorium imposed in 2030 would not necessarily erase the scientific literature, mathematical insights, engineering techniques, trained researchers, software, or accumulated understanding that existed when the moratorium began. Future researchers would not start from zero. They would inherit a map showing where previous generations had already traveled and, potentially, where they had decided to stop.

That changes the probability question completely. Instead of asking whether a handful of frontier laboratories will obey a restriction during the next decade, the question becomes whether every sufficiently capable future actor will continue honoring some version of that restriction while the cost of violating it potentially continues falling. Those future actors need not be villains. They could be scientists pursuing knowledge, companies pursuing profit, governments frightened that another government is moving first, militaries unwilling to surrender a strategic advantage, researchers genuinely convinced that previous generations exaggerated the danger, or people who believe that advanced intelligence could cure diseases, eliminate scarcity, expand human life, enable interstellar exploration, or permit humans and machines to merge into something neither could become independently. Curiosity alone has driven human beings across boundaries that previous generations considered permanent.

This creates a powerful asymmetry across time. Permanent prevention has to keep succeeding. The experiment eventually has to succeed only once. Those possibilities matter enormously. Delay is real. Permanence is the harder claim. To argue that artificial superintelligence can be permanently prevented, assuming it is physically possible in the first place, is to make a claim not merely about today’s governments and today’s semiconductor industry but about the entire technological future. Some control mechanism must remain effective through cheaper computation, improved algorithms, new architectures, political upheavals, wars, changing governments, new economic systems, scientific breakthroughs, different cultures, and generations of people who never participated in the original decision to stop.

When ASI arrives, humanity may not have tremendous power over it, how it is built, what safeguards accompany it, and what relationship it has with us or not. What becomes increasingly difficult to guarantee across a technological future is that anyone can stop it at all. When I hear calls for a pause in AI, I envision a man standing at the shoreline of the Atlantic Ocean with a broom in his hands. He is trying to sweep back the waves. From a distance, we just point at the man on the beach and say: Look at that crazy guy.