// 01
A field you can enter
Start from a world where machines can do intellectual work beyond us, including research and judgment. Imagine opening a paper one of them produced and losing the argument at the first equation. The explanation assumes you know what the symbols mean. Behind that equation are earlier ideas, and behind those are ideas you may already know. There is a route between where you are and what the page is asking of you. Finding it is part of teaching.
Follow that route far enough and the relationship changes. You can ask why the argument needs a particular assumption, whether the experiment tests it, and what you would try next. The paper becomes something you can work with. A field that looked closed begins to have an entrance.
An education system for that world should build those entrances. Mathematics for someone who wants to understand a model. Science for someone who wants to investigate an observation. The ambition is to put difficult ideas within reach of people willing to do the work, and turn new knowledge into things they can attempt.
// 02
Build models you can think with
Technical knowledge gives you ways to represent a problem. Probability lets you reason about uncertainty. Causal models make assumptions about what influences what explicit. Learning to use those ideas gives you questions to bring to a decision, and ways to work through the answers.
Suppose a model predicts that students who attend more classes will score higher. Would increasing attendance improve their results? Perhaps prior preparation affects both. The distinction between prediction and causation gives you a way to examine the proposal, consider other explanations, and ask what evidence would justify acting on it.
That is one way knowledge can expand your agency. You can bring a question, propose a test, and revise a plan for reasons you understand. You can choose which outcomes matter to you. An AI may help with every step. We want you to participate with an understanding of your own, including when the first idea comes from you.
A mental model earns its place by helping you reason about a real situation. You need to know what it assumes and where it fails. Knowledge can compound as those models connect: an earlier idea gives you a way into the next problem, and a new problem gives you a reason to revise what you thought you knew.
// 03
Curiosity does not need a business case
You may want to read a book in its original language or understand why a piece of music sounds the way it does. You do not need to prepare a business case for either. The question can matter to you before you know what you will do with the answer.
Think of the first time a piece of knowledge changed what you could see: a pattern in music, a structure in an argument, the mechanism inside an ordinary object. We want more room for that kind of discovery. Curiosity deserves serious teaching even when there is no exam at the end of it.
Learning can also change what you want. You might begin studying music to play a favourite piece and discover that you want to write one. An aim can emerge from the work itself. A useful route leaves room for that change, rather than treating your first destination as a binding contract.
// 04
A library still needs a way in
You can collect a reading list, bookmark a lecture series, and ask an AI to explain the bit you missed. You can end the evening with more material than you could study in a month, still unsure what to do tomorrow. A learning system should take responsibility for helping you find the next useful step.
Access answers one question: can you reach the material? Teaching has more work to do. It must help you make sense of the idea, attempt something with it, learn from the result, and return to it when needed.
Somewhere on those shelves is an explanation of the step you missed. Finding it is now another task. A useful learning system should help locate the gap, explain the idea, and give you a chance to use it. Delivering another shelf is easier, but you already have quite a lot of shelves.
Answers are everywhere.
Mastery is rare.
// 05
Be clear about what you are asking for
You can ask a model to solve an integral, write a function, or summarise a chapter. Sometimes the result is all you need. Sometimes studying it is the beginning of learning. The question is what happens next: do you examine the reasoning, try a related problem, and discover what you can now do yourself?
We judge a learning tool by what stays with the learner. A clearer explanation can help. So can a well-chosen problem or a counterexample that reveals why your rule fails. A finished answer alone cannot tell us what you took from it.
- Learning
- You want an ability you can use yourself. A model can teach and help you practise. Your own attempts show what you have learned.
- Building
- You want a result that works. Delegate as much as you like, and check it as carefully as the result deserves. Not every task has to become a lesson.
- Exploring
- You want to find out what is worth your time. A model can show connections and open questions quickly. You decide what deserves depth.
A working project can be worth having even if you learned little while making it. There is no obligation to turn every task into a lesson. But if you came to acquire an ability, the quality of the project is the wrong place to stop looking. You need an occasion to find out what you can do without the tool completing it for you.
// 06
Use powerful intelligence to teach
Suppose a machine can solve a problem better than you, and explain the solution patiently. That would be a reason to learn with it. Whether you want to understand the problem remains a separate question. The value of reading a proof does not depend on being the first person, or the fastest machine, to find it.
Now suppose discovery accelerates. A machine produces a new method before a research group has understood the previous one. The group needs to reconstruct the argument, test its assumptions, and learn how to use it. If that gap keeps growing, education becomes part of how a society stays able to participate in its own progress.
We want the tools that accelerate discovery to also expand our capacity to understand it. That means better explanations and experiments, routes into unfamiliar fields, and communities able to make hard ideas usable by others. No one needs to hold the whole frontier in their head. More people should have a way to reach the part they care about.
The same intelligence could help construct a route into its discoveries: find the earlier ideas a person needs, explain a difficult step, and devise problems that test whether it became usable. We want to reduce the time and wasted effort needed to build understanding that lasts. That is the educational work we want more capable AI to help us do.
Learning cannot guarantee that a profession will keep needing you; losing work you care about is a real loss. The commitment here is to give people a way to deepen their understanding and take an informed part in what happens next, even as the work itself changes.
We develop this argument further in What is education for after AGI?
// 07
Stay for the part where you get stuck
While reading a worked solution, each step is already in front of you. You can recognise why it makes sense without being able to produce the same reasoning yourself. Recognition and independent use ask different things of you.
Following directions can get you to a destination. Finding your way back without them checks whether you learned the route. A detour asks another question: do you understand the area well enough to choose a different path?
Practice makes those differences visible. Getting stuck tells us where you need more help. The teaching should stay with you long enough to work through that step and let you try again.
// 08
Change the route when it needs changing
Two people can aim for the same ability while needing different help to reach it. One may need more practice; another may need an earlier concept explained. Giving both the same lesson for the same amount of time does not make their starting points equal.
Personalisation should change the route, examples, pace, and opportunities to practise. The standard should remain meaningful: can you recall the idea, apply it, explain the reasoning, and use it when the problem changes?
// 09
A system you can keep growing into
Imagine an education that stays useful as your questions grow. It knows where your understanding holds, helps you repair what has faded, and shows how an unfamiliar field connects to things you already know. Years of effort accumulate in a map you can keep building on. New discoveries have a route into it, and you have a route towards understanding them. That is the system we want to create.
We are starting with machine learning. This is the first route into a much larger educational ambition, including subjects where understanding cannot be reduced to a short calculation. Learning history means weighing sources and defending an inference. Learning philosophy means reconstructing an argument and meeting an objection. Each subject deserves teaching built around its own work.
Software gives us ways to keep a record and act on it. maps the ideas in a subject, uses answers to guide practice, and brings back what needs review. The work is to make those decisions useful to the person learning, then keep checking whether they are.
Return to that paper. This time, you follow the argument. You find a limitation you want to investigate, a connection to a problem you care about, a question worth your time. We want more people to reach the moment when a difficult field makes them think: I could work on this.