How to learn
You watch the lecture. Every step makes sense. You take notes, you nod, you close the laptop feeling that today was a good day.
A week later you sit down to build the thing, and you cannot write the first line.
Nothing is wrong with you. You did what almost everyone does: you took the feeling of learning for learning. They are different things, and the uncomfortable result from the research is that they often point in opposite directions.
This is what I want every student of mine to know before the first session. The frame is borrowed from Farnam Street’s guide to learning, which is worth reading in full. I have added the studies behind it, and what changes now that a machine can hand you any answer.
The feeling of learning is a bad guide
In 2019, Louis Deslauriers and his colleagues at Harvard ran an experiment in an introductory physics course. Students were randomly assigned to one of two classes covering the same material with the same handouts. One was a polished lecture. In the other, students worked on problems in small groups while the instructor moved between them. Afterwards everyone took the same test, and everyone was asked how much they felt they had learned.
The students in the active class scored higher. They also said they had learned less.
The better method felt worse. The reason is effort. A smooth lecture is easy to follow, and ease feels like understanding. Working on a problem you cannot yet solve feels like confusion, and confusion feels like failing. It is one course at one university, so I would not lean on the exact numbers. But the pattern shows up everywhere researchers have looked. Robert Bjork, who has studied it for decades, calls these desirable difficulties: the conditions that slow you down while you study are often the ones that make the learning last.
The same thing happens when you study alone. In 2013, John Dunlosky and four colleagues reviewed the evidence on ten common study techniques. Rereading and highlighting, the two that students use most, were rated low. They are popular for a reason. The second time you read a page it is familiar, and familiar feels like known.
But recognising something when it is in front of you and producing it when it is not are two separate abilities. Only the second is any use when you are building.
Farnam Street puts it in one line: “The greatest enemy of learning is what you think you know.”
Where learning comes from
There are two sources: what happened to other people, and what happens to you.
Other people’s experience is the cheap one. Someone spent ten years finding out, and you get it in a book, a paper or a course. Take as much of it as you can.
The catch is that facts expire. Samuel Arbesman wrote a whole book on this, The Half-Life of Facts: in every field, a predictable share of what is known today will be overturned or made obsolete, and the rate is measurable. In AI the half-life is short. The framework you learn this quarter may not exist in two years.
So spend your effort on what decays slowly. A tool goes out of date. The problem the tool was built to solve does not. That is why in my classes we derive the idea first and name it last: if you know why a thing has to exist, you can pick up its replacement in an afternoon.
Your own experience is the expensive one, and it does not teach you automatically. It teaches you only if you look at it.
Chris Argyris explained the difference with a thermostat. A thermostat that switches the heating on when the room drops below 20 degrees is learning in a single loop: it sees an error and corrects it. A thermostat that could ask why it is set to 20 at all would be learning in a double loop. It questions the belief that produced the error, not just the error.
Fixing the bug is the single loop. Asking what you believed about the system that made you write the bug is the double loop. The first gets you working code today. The second means you do not write that bug again.
It takes ten minutes. After a project, a failed attempt or a hard session, write down three things: what I expected, what happened, what I will do differently. Most people skip this, and it is why years of experience do not reliably make someone good.
Four methods that work
Explain it without the name
This is the Feynman Technique, and it is the best test of understanding I know.
- Pick one idea.
- Explain it, in writing or out loud, as if to a twelve-year-old. Plain words only.
- Notice where you stall, wave your hands or reach for jargon. That spot is the gap. Go back to the source for that spot alone.
- Explain it again, shorter.
Then do the hard version: explain it without using its name. Not “an API is an interface”, but what problem two programs have when they need to talk, and what would have to exist to solve it. Where you stumble is where the explanation stopped being yours.
Close the book and recall
In 2006, Henry Roediger and Jeffrey Karpicke had students learn a short passage. One group read it four times. Another read it once, then three times wrote down everything they could remember, with no feedback.
Tested five minutes later, the rereaders did better: 83% against 71%. Tested a week later, it had reversed. The rereaders remembered 40%. The group that had practised recalling remembered 61%.
Rereading wins for five minutes, which is exactly long enough to convince you it works.
Pulling something out of memory is what fixes it there. So after every session, close everything and write what you remember on a blank page. Then check. Whatever you missed is your study list.
Come back after you have started to forget
In 1885, Hermann Ebbinghaus memorised lists of nonsense syllables and measured how fast he lost them. Forgetting is steep at first and then levels off. Each time you come back to something just as it is fading, it fades more slowly the next time.
So do not study a thing once for a long time. Study it briefly, several times, with gaps: the next day, a week later, a month later. Three sessions of twenty minutes spread over two weeks will usually beat one hour tonight. Dunlosky’s review rated this and practice testing as the only two techniques of the ten that deserved a high rating.
Cramming works for an exam tomorrow. You do not have an exam tomorrow. You have a career.
Practise at the edge of what you can do
Anders Ericsson spent his working life studying how experts get good, starting with violinists in Berlin. His answer was not hours. It was a particular kind of hour: a task just beyond what you can currently do, a specific goal, full attention, and feedback fast enough to correct you. He called it deliberate practice.
Following a tutorial is not that. You are typing someone else’s decisions. Finish the tutorial, then change something it did not cover. Break it and repair it. Then build the same thing with the tutorial closed. People learn by making things, and the making has to be a little too hard.
What AI changes
All four methods have one thing in common. They make you produce something before you are ready. AI lets you skip exactly that part.
Hamsa Bastani and her colleagues tested this with nearly a thousand high school maths students in Turkey. Some practised with a chatbot much like ordinary ChatGPT. Some practised with a version built to give hints instead of answers. Some had no AI.
In the practice sessions, the students with the ordinary chatbot did 48% better than the students with no AI. Then came an exam without it. They scored 17% worse than the students who had never had it at all. The students with the hint-giving tutor showed almost none of that harm.
Again it is one school, so hold the numbers loosely. The shape is the one we have already seen twice: the method that looked best in the moment taught the least.
So the rule in my classes is attempt first, then bring in the machine. And notice that the same tool is the best study partner you have ever had, if you turn it around. Ask it to quiz you. Ask it to find the hole in your explanation. Ask it to be the twelve-year-old. Use it to ask you questions, and it makes every method above easier to do.
A week of this
- Before the session, write down what you already believe about the topic. Guess. Being wrong on paper is the point.
- Straight after, close everything and write what you remember. Then check it against your notes.
- The next day, explain the main idea out loud without its name.
- Within three days, build something small with it, with the tutorial closed. Bring in AI only after your own attempt.
- A week later, recall it again from a blank page.
- When something breaks, write three lines: what I expected, what happened, what I will change.
None of this is a case for suffering. Difficulty is useful only while it is producing something. If you have been stuck for an hour and learned nothing in the last forty minutes, ask for help.
Nobody can do this for you
Louis L’Amour, who left school at fifteen and read his way to an education, wrote that all education is self-education. I run a school and I think he was right.
A course can give you a sequence, good problems, feedback, and a room of people doing the same hard thing. It cannot recall for you, explain for you or practise for you. Those happen in your head or they do not happen.
If studying feels easy, be suspicious. If it feels hard and you can say what you now know that you did not know yesterday, keep going.
Easy is how it feels when nothing is changing.