LLMs as tutors: one difficult idea at a time
I have been using GPT to work through a difficult idea: four-dimensional spacetime. What interested me was how readily it could break an intimidating subject into a sequence of smaller questions. I could stop at the point where I lost the thread and ask for another explanation.
That is a valuable property in a tutor. A book has to choose its pace in advance. A classroom has to accommodate several people at once. A conversation can stay with the particular step that one person finds difficult, for as long as it remains useful.
My experiment started with familiar things: a moving train, an observer, a clock. From there, the conversation could move towards the harder relationship between space and time. The attraction was the ability to adjust the explanation as I went. A compact definition might be enough for one part; another needed an analogy, a drawing or a slower treatment.
I was impressed by how accessible the interaction made the subject feel. I was also reminded of a problem I encounter in teaching: an explanation can feel clear before the learner has actually understood it.
Let the learner do some of the work
A model can keep producing polished paragraphs indefinitely. After a while, reading them starts to resemble progress. The real test comes when you have to explain the idea yourself or apply it to a slightly different problem.
I would like a tutoring interface to make that moment central. Explain one step, ask the learner a question, then wait. Use the answer to decide where to go next. If something is missing, return to that point with a different example. If the learner can already handle it, move on.
For my spacetime example, this might mean asking what information is needed to describe an event. The learner has to supply both a place and a time. From that answer, the tutor can find out whether the basic idea is understood before adding more demanding material.
A useful prompt would be:
Help me understand this one step at a time. After each step, ask me a question and wait for my answer. If my explanation is incomplete, help me find the missing part before continuing.
This gives the conversation a different rhythm. The learner has to participate. It also produces more information about where the explanation is working and where it is merely sounding convincing.
I would add another instruction: ask where an analogy breaks down. The familiar picture of a heavy ball on a stretched sheet can offer an entry point into curvature, but it is a simplified representation. It leaves things out. A tutor should help the learner notice those limits before the picture becomes a substitute for the concept.
The interface has a job to do
A general chat box asks the learner to design much of the lesson. They must know when to request an example, when to test themselves, and when a fluent answer needs checking. Those are significant demands on someone who has just admitted they do not understand a subject.
There is room for better products here. Let me choose a topic and describe what I already know. Keep track of the concepts we have covered. Give me a small problem that requires me to use them. Bring back an earlier idea when I need it, and show me which sources support the explanation.
Accuracy matters throughout. A convincing sequence of teaching examples can accidentally become a story about how a scientist made a discovery. The examples may be useful while the historical attribution is wrong. Scientific details and historical claims should be checked against reliable material; the model's confidence is a poor guide to either.
My interest is in the quality of the learning interaction. Can I ask a question without embarrassment? Can we slow down? Can I try an explanation, get useful feedback and try again? Those are ordinary requests, yet meeting them consistently would make a substantial difference to how people learn.
The capability is already interesting enough to experiment with. I want an interface that makes the next useful question easier to find, and gives me enough space to answer it.
Revised and expanded from my LinkedIn article of 14 March 2025.