How to learn hard things fast with AI

Nick Saraev made a great video on learning hard things fast with AI. The core of it: asking AI to explain things feels like learning, but mostly isn't, and there are nine other ways to use it that actually work.

It's a long video, so this is the short, drawn version. I'll use one example all the way through: you have three weeks to get ready for a React Native interview. And one picture: a gym.

Does reading it again help?

Not much. You don't learn by taking things in. You learn by getting them back out: recalling, explaining, predicting, trying.

There's a classic study on this. Students either read a text four times, or read it once and then tested themselves on it three times. Five minutes later, the rereaders remembered more. A week later, it flipped hard.

Rereading is watching someone else lift. Testing yourself is lifting. One feels easy and changes nothing. The other feels hard, and that's the part that builds the muscle.

Why does asking AI feel like learning?

Because you read a clear explanation, nod along, and it makes sense. That feeling is real. It just isn't memory.

So the question isn't "should I use AI to learn?" It's which job you give it. Most people only ever hire it as an explainer. That's one job out of ten, and it's the only one where you do none of the lifting.

Think of them as the staff at a good gym. Here's each one, used on the interview.

Before you start

The interviewer: what do you actually need?

A good gym starts with an intake chat, not a workout. What's the goal? What can you already do? When's the competition?

Do the same before you learn anything. Don't ask AI to teach you React Native. Ask it to interview you first:

Before you teach me anything, ask me questions one at a time until you know my goal, what I already know, and how I'll be tested.

In our example, the answers change everything. The interview is in three weeks, it's live coding with a timer, and you already know React for the web. So you don't need "React Native". You need lists, navigation, layout on a phone, and performance, done fast, under pressure. A week of JavaScript basics just disappeared.

Be specific in your answers. "An interview" gets you a vague plan. "A 45-minute live build of one screen, plus questions on performance" gets you a useful one.

The mapmaker: what's the route?

Next, the program: what to train, in what order. Without one, half your energy goes on wondering what to study next.

Map out React Native for a web developer with a mobile interview in three weeks. List the parts in order, what each one depends on, and where people usually get stuck.

You get something like: layout and styling (Flexbox defaults to columns on a phone), then lists, then state and effects, then navigation, then performance. And a warning that most people get stuck on why long lists stutter. Now you know where the hard week is before you hit it.

While you learn

The explainer, with one rule

This is the job everyone already uses, and it's fine. The problem is what happens after.

Say you build a list screen and every row re-renders when you type in a search box. You ask the AI about just that part, at your level, and it explains it. It makes sense. That's not the end. Close the chat and build the whole screen again from a blank file, with no help. That's the rep that counts.

1. 1Try it, get stuckYourows re-render every time you type
2. 2Ask about just that partAIexplained at your level, nothing more
3. 3Start over, no helpYourebuild the screen from a blank file
4. 4Get quizzed on itAIone harder question at a time

The Socratic questioner: do you really get it?

A coach who keeps asking "why?" until you hit the edge of what you know. It never hands you the answer.

Don't explain anything. Ask me questions about how FlatList stays fast until I find the gap myself.

It might ask: what happens to a row when the data changes but its key stays the same? If you can't answer, you just found the gap, and you found it yourself, which is why it sticks.

The listener: can you explain it back?

Teach the move back to the coach. Say it out loud (hold your dictation key and talk), sketch it on paper, then write it in three sentences. Paste all three:

Here's how I think FlatList handles 10,000 rows, said, drawn and written. Grade it against the official docs and tell me exactly what I missed.

Saying, drawing and writing are three different ways of pulling the same thing out of your head. If one of them is shaky, so is your understanding.

When you test yourself

The examiner: where's your ceiling?

A max-out day. Questions get harder until you start guessing.

Quiz me on React Native performance, one question at a time. Make each one harder. Stop when I start guessing, and tell me where my level is.

Passing levels one to four and failing five tells you exactly what to train next.

The checker: is your working right?

Form check. Not "did you lift it?" but "how?". Paste your solution and ask it not to rewrite anything:

Here's my list screen. Find what's wrong or missing, and tell me if there's a simpler way. Don't rewrite it.

It might point out the new arrow function you create in renderItem on every render, which quietly breaks every optimisation you added. Next time, your process is faster, not just this answer.

The sparring partner: can you do it for real?

In a boxing gym, you don't learn to take a punch from a book.

Play a tough senior engineer interviewing me for a React Native role. One question at a time, push back on vague answers, and give me 60 seconds for each.

Say "I'd optimise the list" and it pushes back: how, exactly? By the third round you're saying "memoise the row, keep renderItem stable, and give getItemLayout fixed heights" without thinking. It won't be exactly like the real thing. It doesn't have to be. The first rep is the scariest one, and now it's done.

After the session

The diagnostician: what keeps going wrong?

Same injury, different lifts? There's one cause underneath. Paste your last few sessions:

Here are the mistakes from my last ten sessions. What one misunderstanding do they have in common?

In our example: you keep creating new objects and functions during render, so things you memoised still re-render. The root isn't lists, or effects, or navigation. It's that you don't yet see how React compares props. Fix that once, and three different mistakes go away.

The clerk: keep the log

Every good lifter keeps a training log. The clerk turns your messy notes into a clean outline, and your mistakes into flashcards. All the thinking stays yours; it just tidies it.

Turn my notes into a clean outline and ten flashcards. Don't add anything I didn't write.

Then review the cards on a schedule. Each review pulls the memory back up, and each time it fades more slowly. That's called spaced repetition: a day later, three days later, a week later, two weeks later.

How hard should it be?

You grow fastest with a weight you can just about lift. Too light and nothing happens. Too heavy and you drop it. Psychologists call that sweet spot the zone of proximal development; lifters call it progressive overload.

AI is good at finding your weight:

Quiz me from easy to hard and tell me my level. Then explain this one idea three ways: for a beginner, for someone at my level, and for an expert.

It's also good at finding the one explanation that clicks. Ask it for the three best explanations of a topic for someone at your level, and why each one is good. A great explanation, usually written by a person, can save you a week.

When is AI bad for learning?

When it's a forklift. A forklift moves the weight for you, and the weight gets moved. You just never get stronger.

The struggle is where the learning happens. If AI always gets you unstuck, you skip the part that builds anything. It's the same reason people who follow a GPS everywhere build a weaker map of their own city. Watch for three signs:

• It explained something and you felt you understood it, but you never tried it without help.
• It agreed with you when you were only half right.
• It gave you numbers or sources, and you never checked them.

Prompts to copy

The short version

• You learn by getting things out of your head, not by putting them in. Test yourself more than you reread.
• Explanations feel like learning. After any explanation, start over and do it with no help.
• AI has ten jobs for a learner. The explainer is only one, and it's the one where you do no lifting.
• Use it as a coach that makes you do the reps, not a forklift that does them for you.

The ten jobs and the "redo it without help" rule come from Nick Saraev's video, which is worth watching in full. The study is Roediger and Karpicke, "Test-enhanced learning", 2006. The gym, the interview example and any mistakes are mine.

Related: building with AI, and building AI in.