Amit Sahai, a professor at UCLA and a leading cryptographer, has written a remarkably clear-eyed essay about the structural challenge AI poses to human agency in mathematics and beyond. His argument is not the familiar doom-or-boom binary. It is institutional: as AI systems generate breakthrough ideas faster than any individual can absorb them, humanity needs to massively scale the number of people capable of deep mathematical understanding — not to compete with machines, but to audit them. Sahai's central thought experiment is a one-terawatt nuclear fusion plant designed by an AI using principles no human has conceived. Before construction, someone needs to understand why the design works, how failures propagate, and what the model's assumptions actually are. A theorem only exists within a model. Understanding the guarantee means understanding the model, the experimental evidence for it, and our uncertainties about its accuracy. This is not decorative philosophy — it is an engineering requirement for civilizational risk management. The essay frames this as a collective, not individual, responsibility. Sahai explicitly acknowledges that AI systems may prove mathematical guarantees more reliably than humans can. The issue is not competence but comprehension: decisions of enormous consequence should not rest on reasons no human community understands. He recalls undergraduate peers who left mathematics because they couldn't keep up with the fastest students, and warns the entire research community is about to experience the same vertigo — and must not draw the same conclusion. Sahai proposes what he calls a 'deployable intellectual reserve' — communities of mathematically sophisticated people that humanity can call upon to help understand consequential AI-enabled breakthroughs. This is not a research lab or a think tank. It is closer to an institutional standing army of understanding, ready to spend a term or a year absorbing a single extraordinary set of ideas, with AI assistance. The model is collaborative study, not solitary genius. The counterargument — that AI will make each researcher so productive that fewer are needed — gets a direct rebuttal. Each human is biologically limited in how fast and how deeply they can think. Depth of understanding needs time and a pace of life that humans can sustain. The bottleneck is not computation but comprehension, and comprehension does not parallelize inside a single skull. You scale it by scaling people. Notably, Sahai used GPT 6 Astra to help draft the essay and credits Terence Tao among his reviewers. This is not an anti-AI position. It is a pro-human-agency position that takes AI capability seriously enough to ask what institutional infrastructure must exist alongside it. The essay is short on implementation details — funding mechanisms, institutional design, incentive structures — but the framing is unusually precise for a genre that tends toward hand-wraving. The implicit policy question is enormous: who pays for a civilization-scale expansion of mathematical literacy, and how do you sustain communities of understanding when the economic incentives point toward automation? Sahai does not answer this. But he has named the problem with rare clarity, and the problem is real.
