For most of history, doing intellectual work produced two things: the work, and the worker.
A PhD student spends three months trying to prove something. At the end of those three months, maybe there is a proof. That is the visible output.
But something else happened during those three months. The student tried approaches that failed. They learned which tricks almost work and which ones are dead ends. They stared at the same object long enough to develop some intuition for it. They probably read six papers they otherwise wouldn't have read. Maybe they started noticing a connection that had nothing to do with the original problem.
The university gets a paper.
The student becomes a better scientist.
We tend to count the first one.
I was watching Terence Tao talk about AI and mathematics recently when he made a point that I haven't been able to shake. AI is getting good enough to solve some of the kinds of problems that graduate students traditionally work on. And that creates a slightly strange problem.
Those problems were never just cheap scientific labour. They were how we made scientists.
Tao calls this our "seed corn." If you replace the graduate student with an AI, you may still get the graduate-student-level paper. Maybe you get ten of them. But twenty years later, where does the senior researcher come from?
This seems like a much bigger idea than graduate school.
AI is starting to separate Getting the result From Becoming the kind of person who could have produced the result.
And I'm not sure our idea of productivity knows how to account for that.
A junior programmer debugging some terrible code is ostensibly trying to make the program work. But he is also building a mental model of how software breaks.
An analyst building a financial model is producing the spreadsheet, but she is also developing some instinct for which assumptions actually matter.
You can probably think of dozens of examples like this. Much of what we call junior work has always had this double life. It produces something useful now, while quietly producing human capability for later.
That arrangement was so normal that we barely had a reason to distinguish the two.
Now we do.
Because AI can give us the artifact without necessarily giving us the apprenticeship.
There is another part of Tao's argument that makes this weirder.
Mathematics is beginning to run into what he calls "proof indigestion."
A proof doesn't go straight from nonexistent to human knowledge.
Someone first finds it. Then it has to be checked. Then people have to understand what the actual idea was. Someone has to figure out whether that idea connects to anything important. The ugly first proof gets reorganized. The right abstraction gets extracted. Eventually, if the result matters enough, it gets compressed into the clean version that a student might see years later in a textbook.
AI is rapidly speeding up the beginning of that pipeline: generating possible solutions and checking whether they work.
It isn't speeding up the rest nearly as much.
So mathematicians are starting to face something they didn't really have before: more valid results than they have attention to properly digest. Tao says he has already stopped trying to keep up with everything happening in his own field.
This sounds like an information-overload problem. I think it's more interesting than that.
We are used to thinking of knowledge as something that is either discovered or undiscovered.
Maybe that's becoming the wrong binary.
There is a huge distance between An answer existing And A field understanding it.
Imagine an AI system proves 100,000 new mathematical statements tomorrow and places every proof on a server. Suppose every one of them is correct.
How much more mathematics does humanity know?
Clearly more than yesterday.
But probably not 100,000-theorems-more.
Someone still has to know which five matter. Someone has to notice that theorem 34,281 is really the same phenomenon as theorem 7,204 in disguise. Someone has to realize that one weird proof technique can be generalized. Someone has to throw away the 99,000 results that are technically new but intellectually useless.
The answer is no longer the entire product.
Understanding is part of the product.
And understanding has a much lower bandwidth.
You can think of science as having a write speed and a read speed.
AI is about to make the write speed absurd.
Our read speed is still mostly human.
Those two things together create a fairly uncomfortable possibility.
We could simultaneously have More knowledge that requires expert judgment And weaken the process by which people acquire expert judgment.
AI gives us more proofs to understand, more code to review, more papers to evaluate, more possible experiments to run.
At exactly the same time, it offers to perform the boring work through which a beginner used to develop the ability to do that evaluation.
We could end up with an abundance of answers and a shortage of people who know what to do with them.
That is a much stranger failure mode than "AI takes people's jobs."
And it also explains something about how we should use these systems.
Tao makes another distinction I find useful. Humans and AI aren't necessarily sitting on the same one-dimensional ladder where one is simply "smarter" than the other.
They have different shapes.
A great mathematician might spend months choosing one problem because she suspects that even failing on it will reveal something deep.
An AI can attack a thousand problems.
It can try method A on all of them, then method B, then some obscure technique from a paper written in 1970, then combinations no human had the patience to check. Ninety-five percent of those attempts can be useless and the economics may still work.
Humans have traditionally been very good at depth.
AI is becoming extraordinarily good at breadth.
The mistake would be to use AI merely as a faster imitation of ourselves.
Its interesting use might be to search spaces we simply cannot search, then hand the strange things it finds back to humans who can ask: What the hell is this?
But that only works if those humans still exist.
There is an obvious objection to all of this.
Maybe we're romanticizing struggle.
There is nothing inherently noble about spending four hours debugging a missing semicolon. Students used to calculate square roots by hand too. We gave them calculators and mathematics survived.
In fact, AI could make apprenticeship much better.
A student can have an infinitely patient tutor. They can test an idea in seconds instead of waiting two days for office hours. They can attempt problems far above their current level because the cost of getting stuck has collapsed.
Maybe the next generation learns Faster, not less.
I think that's entirely possible.
So "people need to struggle" is the wrong conclusion.
Most struggle is useless.
The hard question is that We don't actually know which parts of the struggle were doing the teaching.
That is what worries me.
We can measure whether the code works.
It is much harder to measure whether debugging it gave someone taste.
We can measure how many papers a lab publishes.
It is much harder to measure whether a young researcher is developing the intuition required to ask a question nobody has thought to ask.
The danger isn't automation itself. It is optimizing away formative work because its second output was invisible.
And our institutions are particularly vulnerable to this because almost every metric we use measures the artifact.
Papers published. Features shipped. Tickets closed. Models built. Problems solved.
Imagine two research labs.
The first produces 50 papers this year with heavy AI automation.
The second produces 20, but its students leave substantially better at doing research.
Which lab was more productive?
I don't think the answer is obvious anymore.
Tao begins the interview with a nice analogy.
Doing science is like hiking towards a waterfall. You get lost. You make a map. You notice something strange along the way. Maybe you discover another waterfall you weren't looking for.
AI can be a helicopter.
It takes you directly to the waterfall.
Objectively, this is an incredible improvement in transportation.
But you don't know the terrain.
The naive response is to reject the helicopter and keep hiking everywhere.
That would be stupid.
The harder response is figuring out When the point was reaching the waterfall, and when the point was learning the terrain.
I think that distinction is going to show up everywhere.
Sometimes the work is purely instrumental. You need the answer. You don't care about becoming better at producing that class of answer yourself. Automate it.
But sometimes doing the work changes what you are capable of noticing next.
That work has value even when a machine can already produce the immediate output better than you can.
As AI gets better, I suspect one of the most important skills won't be knowing how to use it.
It will be knowing When not to let it think for you.
Not because human thinking is sacred.
Because sometimes the thing you're producing isn't the answer.
It's you.