The Underutilization Hypothesis: On the Allocation of Existing Artificial Intelligence, Human Comparative Advantage, and the Question of What Computation Is For

Public debate about artificial intelligence runs along two tracks: whether deployment is outpacing our ability to govern it, and when systems will match general human capability. These look like opposing camps, but they share an assumption - that the decision variable is the quantity of machine intelligence. One asks for less, while the other asks how soon there will be more. This talk argues that the more consequential variable is allocation. A large fraction of the realizable value of today's AI systems may go unrealized, not because the systems are insufficiently capable, but because capability flows toward applications with appropriable returns and low integration costs, and away from the diffuse, institutionally fragmented problems that make up much of scientific, administrative, and educational work. That pattern may be the predictable behavior of a functioning market applied to a technology with large positive externalities, meaning that capability improvements alone may not correct it. I'll introduce a framework for reasoning about return per unit of compute, a model of where human conceptual work retains economic value without any appeal to human superiority, and what AI's gold-medal Mathematical Olympiad performance does and doesn't establish. Environmental constraints are treated as binding, not as an objection to wave off.