// 01
How much of the possibility can you use?
Imagine asking an AI to build a program that predicts which students will pass a course. It writes the code, tests it, and explains the result. Suppose it does each of these things better than you could. You have acquired a program. Whether you have also learned something about prediction depends on what you do with it.
You could run it and leave the subject there. Or you could ask which information the prediction uses, why the test is convincing, and what would happen if the course changed. The assistant can help answer those questions. Understanding the answers would give you something beyond this particular program: a way to think about the next one.
This changes the question worth asking. Grant that a model becomes better at the work and better at judging it. What would you want to build, investigate, or understand with that capability available? Learning gives you a way to take part in answering. We want to explore how far that participation could go.
// 02
Two different maps
One map shows the knowledge available to you: books, results, explanations, and things a model can produce on request. Another shows what you understand well enough to reason with. The maps overlap, but access to an answer does not put it on both.
A stronger assistant can help you change the second map. It can explain a confusing step, offer a counterexample, or help you test an idea. There is no reason to make learning harder by refusing good teaching. But an explanation being available and a person making sense of it are still separate events.
Imagine using that help to enter a field you had written off as inaccessible. Work backwards from a paper to the mathematics it assumes, study a worked example, then try a problem yourself. Each idea you can use gives the next conversation more substance. That is the possibility we care about: tools powerful enough to help people undertake a more ambitious education.
// 03
What you choose to hand over
Consider a small example. A calculator says that 48 times 21 is 1,008. Knowing that 50 times 20 is 1,000 gives you a rough check. You have delegated the exact multiplication while keeping a sense of the answer's size. That estimate cannot certify every digit, but it can tell you when a result deserves another look.
Return to the student predictor. Imagine that its training data includes each student's final exam result, although the program is supposed to make predictions at the start of term. It can appear accurate by using information it would not have when a real prediction is needed. This is data leakage. The program answers an easier question than the one you meant to ask.
You can ask the AI to audit for leakage, and a capable system may catch it immediately. It can also teach you why the column is a problem. Once you understand that, you can inspect what was known at prediction time, suggest a more appropriate test, and explain the result to someone else. You need not fit the model by hand to participate in those decisions.
Delegation is compatible with understanding. You do not need to reproduce every task you hand over. The useful question is whether you understand enough about the task to decide what you want from it, and when you need help judging the result.
// 04
Learning can change the request
There is a deeper reason to learn than checking an answer. In his essay on learning with AGI, Andy Matuschak describes work in which you discover what you want by trying things. Making an image or designing a product changes your intentions as you see what each attempt does. [1]
The predictor could take that turn. Suppose you discover that a missing assignment was recorded as a zero. Now you want to know whether the student misunderstood the lesson or could not attend it. You began by asking who would pass. You are becoming interested in what help someone needs before the next class.
Predicting an outcome and changing it are different tasks. A model that predicts who will pass does not, by that fact alone, tell you which support would help. Understanding that distinction gives you a new problem to work on: what would let you tell whether an intervention made a difference? You can begin designing the investigation.
This is what we mean by learning for agency. Technical ideas become models you can think with: ways to frame a problem, anticipate consequences, and consider what you could change. They give substance to a question of your own. The model could suggest the question too; understanding lets you pursue it for reasons you can examine and revise.
This is also a reason curiosity need not justify itself in advance. You may study a subject because one small part interests you and discover, through the work, a much larger reason to continue.
// 05
Build the human side of progress
The same distinction matters beyond a single learner. Henry Cohn argues that correct but poorly explained mathematics can leave a community with technical debt: work still needed to understand the result and build on it. [2] Amit Sahai makes a related case for communities of people who can spend time absorbing consequential ideas. His answer includes more people sharing the work, rather than requiring each person to keep up alone. [3]
If discovery accelerates as much as some forecasts imagine, this work could become more pressing. Dario Amodei sketches such a possibility for biology and medicine; it is a forecast, not a settled account of the future. [4] In that scenario, a new result needs a route into human understanding: explanations that reveal its assumptions, experiments that make it tangible, and teaching that connects it to what people already know. AI could help build those routes as well as discover the results.
This does not make learning a race against every new result. It makes choosing what to understand more consequential. Nor does it guarantee a livelihood. If machines take work someone cared about, that is a real loss; telling them to enjoy studying it would not replace the work.
This is the reason we are building as a learning system for a post-AGI world. A new result should have a route into an education: the foundations it assumes, an explanation someone can follow, problems that reveal what they understood, and later checks of what they can still use. The system we want to build connects that work across subjects and years of learning.
Our ambition includes the freedom to initiate: to notice a problem, imagine an alternative, and choose to work on it. Machine learning is our starting point. The connected curriculum, diagnostic, lessons, practice, and review are the first pieces of that system. More capable AI could help us build better teaching and bring new knowledge into reach. The measure of that ambition is what becomes possible for the person learning.
Return to the predictor. You began by asking who would pass, then learned enough to ask what would help someone learn. A different system is now worth building. This is the promise we see in learning alongside powerful AI: the answer can improve, and so can your idea of what to ask for.
// notes
Notes
- Andy Matuschak, What's worth learning if we have AGI?, March 2024. His account of taste connects a repertoire of ideas with the ability to anticipate and perceive the effects of a choice. He also asks whether a sufficiently capable system could learn our taste. The argument here does not require judgment to remain exclusively human. ↩
- Henry Cohn, The technical debt of AI-generated mathematics, September 2026. A guest essay on Terence Tao's blog. Cohn argues that correct but poorly communicated results can create work for the people who must understand and extend them. He also identifies cases where a factual answer is useful before it is fully understood. ↩
- Amit Sahai, We're gonna need a lot more mathematicians, September 2026. A guest essay on Terence Tao's blog. Sahai argues for communities able to absorb and evaluate consequential ideas, without requiring people to repeat every task a machine can perform reliably. ↩
- Dario Amodei, Machines of Loving Grace, October 2024. His forecast of accelerated progress in biology and medicine is explicitly speculative. Our discussion of a possible gap between discovery and human understanding is an argument, not a measured effect or a forecast of how quickly anyone will learn. ↩