I watched a conversation the other day about AI in mathematics research, and one line stayed with me. Terence Tao, the UCLA mathematician and Fields Medalist, described the recent run of results: four years ago, middle-school math, then high school, then Olympiad problems, then a few open questions people had actually tried hard to solve. And then the implication: we may be about to feel very productive while science itself slows down.
The video is here. My notes below mix what Tao argued with what I think he got half right.
The mechanism that worried me
The grad student example is Tao's clearest argument. The first problems a PhD student solves are deliberately small, training wheels to build the taste that later finds real results. Models can now produce work at that level on demand. Replace those students with agents and the papers still appear. The researchers who would have existed ten years later do not. Nothing looks wrong from the outside for about a decade, because output is the only thing most people can count.
The second case is coding. Experienced engineers report writing code five, ten, even a hundred times faster with agents, and also report losing the ability to code by hand, until some can no longer review what the agent produced.
The part we have always skipped
Humans offload. It is what we do. Calculators ended mental arithmetic as a research skill and mathematics survived. Compilers ended hand-written assembly and programming grew. Every time, the layer that went away was the mechanical one, and the human kept the part that required judgment.
The question is which layer we are skipping now. A calculator never replaced the step where you understand why the proof works. An agent can. Tao's own account of the wins is breadth over depth: point a model at a thousand open problems and it clears a few percent. The wins sometimes come from connecting two literatures no human had time to read together, which is how I read the unit distance result he mentioned. Those answers are real. What is missing is a human who absorbed the reasoning behind them, who could stand on that result and ask the next question. We are collecting answers faster than we are building people who can use them.
The Copernicus problem
His best historical example is that Copernicus's heliocentric model predicted planetary positions worse than the refined geocentric system. Kepler's ellipses made heliocentrism actually beat the old model on the data. Now imagine both of them with a model that grades itself by immediate fit. The correct framework gets discarded at first, because it looked worse. Progress in science is not only answers scored against data. It is also a community deciding which questions are worth carrying forward, and that decision is slow, social, and easy to break with a fast oracle.
The black box around the wins
Tao's quietest caveat may be the most important one. The results are real, but we cannot size them. Private companies announce the wins and not the bill: was the solution a hundred thousand dollars of compute or a million? Did the model look at the one problem it solved, or at a hundred problems and clear one? What was the failure rate on the rest?
Without the denominator, "AI solved an open problem" and "AI solved an open problem after ten thousand tries" are the same headline. Those are wildly different achievements, and right now only the company knows which one happened. Tao said it plainly: it is unclear how replicable any of this is. A field learns by repeating what it reports. This part of the progress is structurally un-repeatable from the outside.
Editorial
The line I want to hold
I use agents all day, so this is not a plea to stop. It is a note to myself about which step not to skip. When an agent writes code I did not plan, I read it as if I own it, because I do. When it solves something, I ask for the reasoning before the answer, even though the answer is the part that saves time. And when it produces output at the level of my own training wheels, that is the moment to go slowest, because nobody else is going to teach me what that step taught.
Throughput feels like progress. It is the cheapest thing to measure and the easiest to mistake for the point. The frontier only moves when someone understands why the last answer worked. That understanding is the part nobody can outsource for you, and right now it is the part nobody can even see you losing.