This ran as a guest post on Terence Tao’s blog, written by Amit Sahai, a computer science and mathematics professor at UCLA. (His own note says the draft was converted between file formats with AI help, and credits a model for helping him write it.) Sahai’s argument is not that AI does arithmetic quickly. It is about who will be able to understand what the machines produce.

He opens with the students who left. Undergraduates who could follow the hard ideas perfectly well — just not at the pace of the fastest people in the room — and who mostly gave up on research mathematics. His framing is that the profession is “now entering a time for humility”: a time when everyone in it will know what it feels like to be unable to keep up.

  • The AI systems he has worked with are already producing “beautiful new ideas,” he writes, and doing more than fast calculation or arguments a strong human researcher could already have made.
  • The failure mode he fears is the one his classmates chose: if the machines move faster, find something else to do. Abandoning the work of understanding would be, in his words, “a profound abdication of our responsibility to humanity.”
  • His counter-proposal is a “deployable intellectual reserve” — many research groups, each with sustained support, each spending a term or a year trying to understand an extraordinary set of ideas an AI system produced, with AI systems helping them do it.
  • That is where the title comes from. The enterprise implies a large expansion in the number of mathematically sophisticated researchers worldwide, because major results will arrive faster than any one group can absorb them.

The case for paying for this is a fusion plant.

Imagine an AI proposing a terawatt design that sustains and controls fusion using principles no human has conceived of, with robots ready to build it. Before anyone approved construction, you would want people who understand why the design works: how failures are contained, what happens to the energy stored in the system when it shuts down, how the materials actually behave under conditions nobody has tested. The novelty that makes the design exciting is exactly what removes the option of inheriting confidence from decades of similar plants.

Two concessions keep the piece honest. Human involvement does not automatically improve a technical decision — he links a Nature paper on that — and a theorem only exists inside a model, so understanding the guarantee means understanding the model, the evidence for it, and the uncertainty about both. A footnote carries the political edge: that understanding “cannot belong only to the organization proposing the technology.” A public hearing where the company’s experts are the only people able to follow the argument is not oversight.

The ask is narrow and monetary: staff and fund human comprehension of machine output, not as a brake on the technology but as a condition of using it with agency. He anticipates the objection that AI makes each person so much more effective that fewer people could do this work, and answers it with biology — depth of understanding needs time and a pace of life humans can sustain.

The 253-comment thread on Hacker News

What the thread adds

  • pyridines — reports from AI-assisted coding: “I pored over every single line of code Claude generated with razor sharp scrutiny. I would usually catch issues with every response. I’m catching fewer problems these days.” Their worry is the generalisation of that curve: audit the first fusion plant closely and find nothing wrong; what happens by the ten-thousandth? The anecdote leaves open whether the model improved, the checking weakened, or both.
  • liampulles — the same failure in software terms. Colleagues who hand work wholesale to Claude come back with “classic XY-Problem shit, poor user experiences, and over-complex solutions,” which they read as the argument for domain understanding rather than against it. Their framing: “there are more useful outputs to solving a problem than just a mere solution.”
  • Four commenters — asprs, zx8080, r_lee, d2kx — spend their comments on attribution rather than argument. asprs: “50% of people mistake this guest post for a post by Tao himself… The post is by Amit Sahai.” zx8080 puts it as “Not Terrence Tao post. Beware.”; r_lee asks whether most recent posts on that blog “are not actually Tao’s”; d2kx corrects another commenter making the same mistake. asprs goes further and reads the guest-post pattern as a coordinated industry campaign. That is their theory, offered without evidence.
  • Animats — makes the human-ceiling claim: “Humans are close to their ceiling. AIs are just getting started.” xanderlewis answers: “The whole point of mathematics is to vastly exceed that natural ceiling by gradually building a framework for understanding.” A factual correction matters here, too. kens challenges Animats’s 5,000-engineer Pentium Pro example. Chief architect Bob Colwell’s own account describes over 400 design and validation engineers on the project, distinguishing it from a division that later grew to thousands.
  • avianlyric — argues that engineering already extracts value from systems understood through partial models. itsalwaysgood makes the everyday version: you never built your own power plant either. brap’s reply narrows the objection: earlier systems still had people who understood their models deeply; the proposed future might have none. Individual ignorance and the absence of any human expertise pose different oversight problems.
  • askjdfksdbfhk — defends human understanding against pyridines’s objection: “writing code is a means to an end; doing mathematics research is not, but is the end in itself.” They argue that increasing human understanding is the purpose of pure mathematics. The claim that the essay merely asserts human agency as a value belongs to pyridines, whose comment they are answering.

Where the thread pushes back

robotpepi rejects treating future research capabilities as settled: “We can’t destroy the scientific ecosystem based on vague speculations.” a-dub asks whether machine discovery will plateau as human incentives to share research change. These are objections to the forecast itself, alongside the debate about how to respond if it comes true.

The question the thread kept asking

Who pays for the absorbing? goy writes: “The thing that worries us (the generation trying to enter the job market now) is just our livelihood. Food, rent & accomodation while I do the work.” Their question is explicit: “Who’s gonna pay me, and for what ?” The direct reply from derangedHorse says it is a question “you’ll have to answer for yourself”; elsewhere, that commenter proposes taxing capital returns to expand welfare. That is a distribution proposal, still short of a funding mechanism for Sahai’s research reserve. The essay calls for sustained support without specifying who supplies it.

On reading comments as evidence: HN handles are pseudonymous, the site publishes no per-comment scores, and the ordering here is HN’s own ranking rather than a vote. This is a slice of the thread, not a consensus, and the claims above are attributed to the people who made them.