The answer was never the valuable part
Twenty-seven Fields Medalists signed a declaration this month about AI in mathematics.
Their objection is not about capability. It is that there is a severe misalignment between what AI labs optimize for and what mathematics actually needs, and the loss happens both before and after the proof is finished.
Before it, working on the problem is the training, and training mathematicians is essential for the field to advance and even survive. The declaration calls students and ideas “the most precious resources of our profession.” Years of training have served “not only to produce a final answer or product, but also to develop understanding and the ability to formulate new questions and ideas,” and supervisors assign problems to build exactly that. Hand the problem to a model and the answer arrives, but nobody has been formed by reaching it. A field that stops producing people who can ask the next question does not survive its current generation, no matter what gets proved along the way.
After it, a proof gets argued over, simplified, generalized into a method, taught, and absorbed until every mathematician in the field can assume it. The declaration calls that the human transmission chain, and notes that solutions announced in a rush leave no time for a proper writeup or for isolating the new methods. Terence Tao, one of the signatories, has separately described open problems being non-renewably mined.
I’m not a mathematician. I’ve worked on knowledge management problems several times in my career and I see parallels.
A company has nothing equivalent
Somebody asks whether a customer is still under the old pricing. An agent reads the mail, the contract and the thread, and comes back with an answer. The answer is correct. Three weeks later someone else asks the same question, and the whole thing runs again from scratch over a corpus that has changed in the meantime. This time it may retrieve different material and reach a different answer. Nobody recorded the first one, so there is nothing to compare it against.
And nobody did anything wrong. Each interaction looks like a success from the inside. That’s the trouble: the second one may have given somebody an answer that was out of date. The company knows exactly as much after the second question as it did before the first. What got produced was an answer, and an answer gets consumed and disappears.
The mathematicians’ claim is that a field advances when a result is absorbed, and that speed of production is a different quantity from growth in what the field knows. The same is true of a company, except that mathematics has institutions built for transmitting a result and a company has none. No seminars, no referees, no graduate students working through the result until they can teach it. In a company, transmission is somebody writing the conclusion down where the next person will find it.
Where the analogy breaks down
The declaration is arguing about research mathematics, and two parts of it don’t transfer.
A lot of their concern is about people. Mentorship, credit, the way a young mathematician earns a reputation by doing the work that AI would now do instantly. When they say knowledge transmission they mean a human tradition, passed between people over careers, and I am talking about something far more mundane: whether a fact about a customer is anywhere a colleague can reach it next month.
Mathematics also has machinery a company doesn’t: journals, citation, peer review, a canon. Transmitting a result so others can absorb it is somebody’s job there, badly funded and slow, but it exists. In a fifty-person company it’s nobody’s job. Nothing reviews or circulates the fact that a customer has pushed their renewal to March.
Which is why I think the loss of knowledge is worse in a company than in mathematics. The mathematicians are worried about a working system being outpaced. Most companies never had the system.
Measurement transfers
The part of their argument that transfers exactly is the measurement. If you judge an AI deployment by how many questions it answered and how fast, you’ll get a number that goes up regardless of whether your company knows more than it did last quarter. Those are two different quantities and only one of them compounds.
They’re right that absorption is critical and that AI isn’t optimizing for it. Mathematics has accepted methods of transmitting and absorbing knowledge. Most companies are trying to build theirs from a standing start, while the tools get very good at producing answers that nobody records.
We build a product that gives a company a brain, a set of pages holding what it knows, filled from the messages it already has, with workflows that write back to it as they run and a person approving what gets written. When an agent works something out, the conclusion lands somewhere durable instead of evaporating. That’s the whole idea.