Eero Alvar's AI learning harness: probe, teach, practice

Key ideas

• The biggest risk in learning with AI isn't a wrong answer from the model. It's skipping the struggle and feeling like you learned.
• Following an explanation and having the idea come to mind when you need it are two different skills.
• A new idea starts out tied to the situation it was taught in. Practice widens the set of situations that bring it to mind.
• A lesson can only be fitted to you if the tutor sees how you reason, so you type your raw thinking at every step, not just your answer.
• Each step is small enough to feel obviously true: you predict the next move, it explains, then it checks you got it.
• Good practice is recall that succeeds with effort: hard enough to make you work, not so hard you stall.

Eero Alvar’s rule for learning with AI is to keep the struggle and hand off everything around it. Alvar’s open-source desktop app has an AI agent find the edge of what you already understand, teach one reasoning step at a time, then set practice problems that make you recall the right idea in less familiar settings. It depends on you typing or dictating your raw thinking at every step, so the agent sees how you reason and not just whether you were right. These are my notes on the 23-minute video, with the method laid out so you can try it in any AI chat.

The goal

The aim is the most real, usable understanding for each unit of mental effort. That can’t mean the least effort, because working through the material is where most of the learning happens. So the app doesn’t try to do your thinking. It tries to move your effort onto the material itself and off everything else: working out what to study next, finding the gap, writing practice problems.

Understanding versus using it

You can follow every step of a lecture, really understand it, and still not think of the idea when a problem calls for it later. In the video’s framing, that’s recall: the right facts, connections or moves coming to mind in the right situation. Recall runs on cues. A freshly taught idea only comes back reliably in the setting it was taught in, so it starts with a narrow, weak set of cues, and practice is what widens and strengthens them.

That splits learning into two jobs. Teaching puts understanding in. Practice makes it usable.

Why it watches your thinking

There’s no best lesson in the abstract. A step is redundant only if you already know it, and out of reach only if you’re missing what it rests on. So before anything can be fitted to you, the tutor needs a detailed picture of how you reason, and a right or wrong answer doesn’t give it one: two people can pick the same wrong answer for different reasons. Hence the thinking field on nearly every question, and the confidence slider beside it.

It’s also why the app avoids long explanations. Alvar’s reason isn’t that walls of text are tiring. It’s that every stretch without your input is a stretch where the tutor can’t see you.

Teaching

A teaching session runs in three phases.

1. Probe. Quick multiple-choice questions search for your edge on everything the lesson depends on. In the demo it found the edge within a few questions, which Alvar credits to the graph of earlier sessions.
2. Plan. The agent works out the smallest dependency map from what you know to what you want to understand, and you accept it before teaching starts.
3. Teach. It walks the map one reasoning step at a time, so each step feels obviously true, until the goal itself does.

Every step has the same shape. You predict the next move first, then it explains, then it checks your understanding. With two questions and three thinking fields per step, including one for thoughts you have while reading, mistakes surface right away and get fixed before anything is built on them. Four helpers sit behind the tutor: one researches facts, one reads PDFs, one draws diagrams and one builds interactive demos.

Practice

A practice session is a stream of problems that adapt as you go. The agent’s job isn’t framed as writing problems but as making you recall the right thing from a given set of cues. Difficulty moves in two ways: how many ideas a problem combines, and how familiar the cues are. You send your thinking in pieces, and the agent works out whether you’ve solved it without you writing out a final answer. When you’re stuck it gives the smallest nudge that points the right way, so you still find the answer yourself.

The graph of what you’ve learned

After teaching, the app distills the session into short claims and the reasoning that connects them, and adds them to one growing graph. Arrows show which ideas lead to which, so you can lay the graph out with the deepest reasoning at the top. Sessions link to each other, and you can import your own notes. Practice sessions can draw problems from chains you pick in the graph, and each problem is tagged with the ideas it tests.

The graph shows what you’ve been taught, not how well you know it. Alvar says feeding practice results back into it is the next step.

How to try it

The app is free and runs on your own model keys. The method also works in an ordinary AI chat if you set it up:

• Ask it to quiz you before teaching, one question at a time, to find where your understanding stops.
• Ask for a plan from what you know to your goal, and agree it before the lesson starts.
• Have it teach one reasoning step per turn: you predict the step, it explains, then it checks you with a question.
• Type your thinking with every answer, including how sure you are and what confused you.
• Later, ask for practice problems that need the idea without naming it, with hints only when you ask.

Key terms

Agent harness

Everything around a model that lets it explore and act on some system. In a coding harness that's a computer; here it's your brain.

Thinking field

A box next to almost every question where you type or dictate whatever you're thinking, unfiltered. It's the tutor's main evidence.

Probe

The opening run of quick multiple-choice questions that searches for the edge of what you already know.

Cue

A feature of a situation that brings a piece of understanding to mind.

Confidence

Your estimate of the chance your answer is right, set on a slider with each answer.

Reasoning chain

A record of which ideas together lead to which conclusion, saved from each lesson into a graph that later sessions read.

Quotes from the talk