Start learning

Start here

How to tell whether you actually understand something

You can read a clear explanation, feel it land, and be unable to reproduce a line of it a week later. Here are five ways to find that out today instead.

Fluency is a liar

Read a good explanation of gradient descent and you'll feel you understand gradient descent. Nothing snagged, you followed every sentence. That feeling is fluency, and it comes from the writing being clear, not from anything having happened in your memory.

Which is why re-reading is such a popular study method and such a bad one. Each pass feels better because the words are more familiar, and you're measuring the smoothness of the text rather than the state of your memory.

Two curves over the week after reading. The sense that you know it stays almost flat, while what you can actually write down from memory falls away, leaving a wide gap by day seven. what it feels like you know what you can produce the gap just read it a day later a week later how well you know it
Re-reading tops up the top line and leaves the bottom one alone. The tests below measure the bottom one.

Five tests, quickest first

1. The blank page

Close everything and write down what you just learned, from memory, in your own words. Then compare. Sixty seconds, and most people who felt confident find a list of keywords with the connective tissue missing. The connective tissue was the lesson.

2. The prediction test

The strongest of the five. Take a situation you have not been shown and say what will happen before you check. You can memorize "overfitting is when a model learns the training data too closely" and still not be able to say which of two experiments will overfit.

Sharper version

Predict the failure. "How would I break this?" is a much harder question than "what does this do?", and answering it needs a mechanism in your head rather than a description.

3. Explain it to a twelve-year-old

No jargon: every term you drop has to be replaced by what it means. Jargon is compression, and compression is where the gaps hide. "It uses backpropagation to minimize the loss" can be said with no understanding at all. "It checks how wrong it was, works out which of its settings pushed it wrong, and nudges those" can't. Where you stall is where your understanding stops.

4. Why this, and not the obvious alternative?

What simpler thing did this replace, and what went wrong with it? If you can't say what breaks without a technique, you've learned a name attached to a procedure — and you'll miss the situations where it's unnecessary or the wrong tool entirely.

5. The one-week test

Come back in a week and run the blank page again, cold. Everything you can do five minutes after reading is short-term memory doing the work. A week later that's gone and what's left is what you learned.

A staircase of five rising steps: recognize, recall, explain, predict, and predict the failure. Each step is a harder question you can answer about the same idea. Recognize Recall Explain Predict Predict the failure easy to fake impossible to fake
Five questions about one idea. Reading gets you the bottom step; the top step is the one that shows up at work.

What the results mean

  • Blank page fine, prediction not — you have the description without the mechanism. Go back for why it works, not what it's called.
  • Prediction fine, twelve-year-old not — you probably do understand it and are leaning on jargon. Worth fixing: teaching an idea is how you find the last gaps in it.
  • Everything fine today, nothing after a week — normal, and a memory problem, not a comprehension one. It has a different fix: spaced retrieval.

Why the course works this way

All five tests make you produce something before you're shown the answer. That's why Toorova puts the question before the teaching: a guess is a blank-page test you can't skip, and it's free to get wrong.

It's also why a wrong answer is told why it was tempting rather than just "incorrect". Naming the belief you were carrying is worth more than restating the right answer.

Next