An AI tutor is not a chat window

A general AI assistant is a genuinely good explainer. Ask it what a gradient is and you will get a clear answer, probably a better one than the textbook you were reading. We use one every day, and nothing below is an argument that you should stop.

But explaining and teaching are different jobs, and the gap between them is where most self-taught study quietly fails. Here is the gap, as concretely as we can put it.

A chat answers the question you asked

That is the whole contract, and it is the right contract for getting unstuck. The trouble is that choosing what to ask is most of the skill you are trying to acquire. A beginner asks about the thing that looks confusing, which is rarely the thing that is actually load-bearing. You can spend an afternoon asking excellent questions about matrix multiplication and never find out that the reason the next paper will defeat you is that you never pinned down what a basis is.

A course exists to make that choice on your behalf, in an order, up to a stated outcome.

A chat has no record of how you did

It can recall what you said, in that conversation. What it has no notion of is whether you got it right, because nothing it asked you was marked.

This is not a small difference. Everything useful that a tutor does downstream depends on having that record: spending the next lesson on the thing you missed, dropping the thing you clearly have, noticing that the same misunderstanding has now shown up three times in different clothes. Without marked answers, "adaptive" means nothing more than "it will rephrase if you say you are confused".

A chat lets you off doing the work

The honest failure mode of studying with an assistant is that reading a good explanation feels almost identical to understanding it, and nothing in the session ever tests the difference. You finish the evening with a transcript of lucid paragraphs and no evidence either way.

Practice is the part people skip, and it is skipped because asking a chat to set you problems means grading your own homework. You answer, it agrees, and a model asked whether a confident answer is right will tend to say yes.

What we do instead

• The course is planned to an outcome, from a placement that finds out what you can already do.
• Every lesson carries practice, in six forms depending on the subject: a number, a program, a line-by-line derivation, an argument, an explanation in your own words, and a choice.
• The marking is done by instruments wherever an instrument exists. Programs are run against tests, and a failure names the one test that failed. Working is put through a symbolic check, line by line, so you are told which line went wrong rather than that the answer is wrong. How your work gets marked goes through this properly, including the cases where no instrument can reach and we say so.
• Any sentence of a lesson can still be opened and asked about, which is the good part of the chat window, kept.

When a chat window is the right tool

Plenty of the time. If you are unblocking one specific thing, debugging, or want a second explanation of a paragraph you have already read, open the chat. It is faster and it is free.

Reach for a course when the subject is big enough to need an order, when you keep starting and not finishing, or when you need to be able to do the thing rather than recognise it.

The first course is free from beginning to end, with no card and no clock. If it turns out a chat window was enough for you, that is a fine outcome and it cost you nothing.