Essay

An AI Tutor Needs to Know What the Student Understands

A useful tutor checks what the student can do and chooses an explanation, hint, or exercise from that evidence.

On this page
Mentors and learners crossing a dark field toward a yellow sunset in a textured painted landscape

A student can understand an explanation while reading it and still be unable to solve the problem alone. An AI tutor needs a way to check that difference. Giving a correct answer does not show what the student has learned.

The next step should depend on where the student is stuck. Someone missing a basic concept needs an explanation. Someone who knows the concept but has made a mistake may only need a hint.

Find out what the student understands

In a conversation with Dwarkesh Patel, Andrej Karpathy describes learning Korean online, in a class, and with a one-on-one tutor. His tutor quickly worked out what he already knew and chose lessons that built on it. A book or language model can contain the same subject knowledge without knowing which part a particular student needs next.

A programming example makes the difference clear. A recursive function calls itself to solve a smaller part of a problem. It also needs a stopping condition, often called a base case, where it returns without making another call. Without one it can reach, the calls keep going until something fails.

Two students can make that same mistake for different reasons. One may not know that recursion needs a base case. The other may have written one but changed the input in a way that never reaches it.

The first student needs the stopping condition explained. The second could trace a few calls and check how the input changes. A student who cannot yet follow an ordinary function call may need to practice that first. Giving all three students the same corrected code would hide the difference between their difficulties.

Karpathy is cautious about how well chatbots can work out what a student understands. Refusing to give an answer is not enough. A tutor that keeps saying "try again" may be withholding the explanation the student needs.

Leave the useful decisions to the student

Karpathy makes a related point about nanochat, his project for building a small language model. He suggests using the finished code as a reference while rebuilding it yourself. Writing the code requires you to decide what a function should remember, which operation should happen first, and how to keep the data consistent. Those decisions are easy to pass over when reading someone else's solution.

An AI can supply the code, tests, and explanation together. That may save time, but it also removes decisions the student could have practiced. Asking the student to predict a result or write the next step gives the tutor something to check.

The aim should be useful practice. Spending forty minutes looking for an obscure software setting may teach very little about the subject. A tutor should use the student's attempt to choose a helpful hint, then explain the missing step if the student still cannot make progress.

This approach also needs to match the task. Someone fixing a build before a meeting may just need a working solution. The user should be able to choose tutoring when learning is part of the goal.

Check what the student can do later

One way to check your own learning is to choose two small topics from the same subject. Use unrestricted AI help for one and a tutoring instruction for the other:

text
Ask me to predict or attempt the next step before showing the solution.
If I get stuck, give a small hint or explain the concept I am missing.
After we finish, ask me to rebuild the result or change part of it.

Spend the same amount of time on each topic. Forty-eight hours later, try both again without AI. Start from a blank page, explain why the main pieces are needed, and change one important condition.

This can help you notice what you retained, but it is not a controlled comparison. One topic may be harder, you may already know more about one of them, and the first session may help with the second. It cannot establish which method works better for students in general.

My guess is that unrestricted help will produce a finished result sooner, while well-chosen hints will help more with the later attempt. That is a prediction, not a result from this experiment. Poor hints could waste time, and a direct explanation could teach something useful. The later attempt gives you a way to check what you can still explain and do without the answer in front of you.