A guest post on Terence Tao's blog mocks AI math as a Library of Babel

A narrator who prefers the pile to the proof

The post, hosted on Terence Tao's blog, is a guest piece by Rikab Gambhir, a theoretical particle physicist, and it is satire. Borges's "Library of Babel" (1941) contains every possible book, correct proofs included, and by 2026 standards would be the most productive research institution ever. Gambhir's narrator accepts that standard. Output is judged by how correct it looks, checking is outsourced to the mathematics community, and "It is inefficient to explain or check a result when those resources can be spent generating another."

The facts behind the joke: OpenAI released 722 mathematics write-ups in early October 2026, and the physics community reportedly produced about 2,500 in a recent period. The essay cites a preprint on misleading Lean certificates, and its last lines turn on the reader: "Remember why you chose to go into science and math in the first place!"

The Library of Babel simulator objection, which the essay already made

A commenter with 9k karma opened with "Such a weird take" and sent readers to a Library of Babel simulator:

"It's always gibberish. The odds of getting something that isn't gibberish are astronomically low, like winning the lottery… Yet with AI, we're getting an abundance of very credible -- and, as Tao himself notes, frequently correct -- mathematical proofs." — commenter with 9k karma who ran the Library of Babel simulator

The answer came from a reply that quotes the essay's own line back:

"Rather than generating every possible mathematical paper, they’ve found a way to probabilistically generate things that look correct." — reply quoting the essay's own line

So the disagreement is smaller than it looks. Both sides accept that the machine's output is filtered. The essay's complaint is about what the filter selects for: results that look right to a reader who has no time to check them.

Trust without understanding

A commenter who praised the piece as "well written, deep, funny, and sad" said it does not go far enough. Their worry is the day AI explains itself better than humans do: formal verification, they argued, is not enough to trust a proof unless people can understand it, and "human understandability is the only thing that matters".

A commenter with 11k karma took the opposite side of the logic:

"The whole point of formal verification is to remove the necessity of understanding most of the work." — commenter with 11k karma replying on formal verification

Their example was a claim about Navier–Stokes: they are confident that the equations can blow up because they trust the Lean proof checker and the people who verified the problem statement and result, "despite the actual proof not being human verified yet (as far as I know)". Another reply doubted the premise itself: explaining something well to a human takes "a degree of empathy and sophisticated theory-of-mind that LLMs have not yet exhibited". And a third offered a test that sidesteps understanding altogether: ask the machine to build something functional that we cannot.

Why society funds pure mathematics at all

The founder of Parse (YC S2011), who now builds a theorem prover with built-in AI, called "Why are we even doing all this?" the right question: "we fund pure mathematics because we believe that some pure mathematics will end up being useful." Turing was doing "pure mathematics" when he invented the Turing machine, the founder noted, so the real debate is which is more useful in 2026, the pure mathematics AI does or the pure mathematics humans do.

A reply from a commenter with 11k karma said the essay is about humans, "not usefulness or efficiency", and that putting the latter above the former misses the point. The founder's question and the reply do not meet. One asks whom an institution should pay; the other asks what the paying is for.

Status versus insight

A commenter thought the author was "hiding the ball": humans have long attached status to cognitive insight, theorists over experimentalists, so the grief is partly about esteem shifting once insight can be automated. A commenter with 12k karma disputed the history:

"AI is such a total problem to this mentality, because it disrupts not only insight production but the mechanism that values insight production." — commenter with 12k karma replying on status

Their argument was that insights were only ever valued when they could be communicated well enough for others to use. A software engineer working in DSP and ML gave that point a practical edge from the other side: many human-written papers in the field were never well written or easy to follow, and now an LLM can walk through the algorithm step by step and take questions.

Read next to the HN thread on Derek Thompson's essay about mathematicians mourning, the two pieces cover different halves. Thompson asks how it feels to lose a calling. The satire and its replies ask who verifies the result, and who pays for it.