Amit Sahai says AI discoveries will need far more human mathematicians
Victoria Shinder2 min readAmit Sahai, the UCLA cryptographer, has published a guest essay on Terence Tao's blog under a deliberately provocative title: "We're gonna need a lot more mathematicians". His premise is that the AI systems he works with are already producing new ideas, not just fast calculation, and that research mathematicians are about to learn what it feels like to be unable to keep up. The argument runs against the common assumption that faster machines mean fewer people.
Sahai's case is that understanding does not scale with compute. He imagines many research groups, each funded for a term or a year, working to understand a single set of AI-produced ideas with AI assistance - and calls that possibly the most important mathematical work of the coming years. His concrete illustration is a hypothetical AI design for a one-terawatt fusion plant: before building it, he argues, communities of humans would need to understand why the design works and what justifies confidence in its safety, because a guarantee only holds inside a model, and someone has to judge the model. He proposes the term "deployable intellectual reserve" for mathematically trained people society can call on for exactly that.
Two details matter for how the piece should be read. A footnote insists the understanding cannot live only inside the company proposing the technology - at a public hearing, independent experts have to be able to follow the argument. And Sahai discloses that GPT 6 Astra helped him draft the note, while Tao notes the post was converted from another format with AI. The comments are sceptical: several readers ask who will fund the extra positions when tenure-track jobs are already scarce.
What it means
For software engineers the essay is closer to home than its subject suggests. The same gap is opening in code review: generating a change is getting cheaper faster than understanding one, and a team that measures only output will conclude it needs fewer reviewers at the moment it needs more. Sahai's footnote about independent expertise maps directly onto vendor-supplied AI tooling, where the only people able to explain a model's behaviour often work for the vendor.
The essay offers no mechanism, and critics are right that it is light on funding or policy. Its useful contribution is a reframing: if AI raises the rate at which consequential ideas arrive, the scarce resource becomes people who can check them, and organisations should budget for that capacity explicitly rather than assuming the tool supplies it.