In 1865 William Stanley Jevons, twenty-nine years old, published The Coal Question. Britain was starting to worry about its coal running out, and plenty of people expected better engines to fix the problem. Jevons argued the opposite. Every improvement to the steam engine since James Watt had made coal cheaper to use, so industry found new jobs for it and total consumption rose.
Economists call this the rebound effect. When use grows past the savings, they call it the Jevons paradox.
Cheaper tokens, more tokens
Running a language model at a given quality has become far cheaper since 2023. Spending on inference has gone up anyway, because cheap calls invite agents that make thousands of them where a person once asked one question.
Oversight follows the same arithmetic. In one of my simulations the cost of monitoring a portfolio of agents grew faster than the number of agents, with an exponent of about 1.47. You can play with the curve in the monitoring cost explorer. Cheaper agents mean more agents, and each one adds more review work than the one before.
Jevons got his own forecast wrong. He expected British growth to stall as coal grew scarce, and he never saw oil coming. The mechanism he described has held for a hundred and sixty years, through electricity and computing, and it is holding for inference.