# Learning How to Learn

> Why learning is the skill to build first, the order to work through the books and videos that teach it, and how to run the loop with AI.

_Source: https://leovanschaik.xyz/guides/learning-how-to-learn/_

---

Learning how to learn compounds. That makes it the highest-return skill to build
first. Most skills are linear. You trade hours for ability, and the next skill
starts from zero. This one carries forward. Each new skill costs less than the last.
The fifth language is faster than the first. A new technical field takes a week to
map instead of six months. You start from method, not from scratch.

The belief that you are not a math person, or a language person, or a technical
person is the main thing holding most people back. It is false. Learning changes the
brain physically. Treat any hard subject as learnable on a fixed schedule and it
becomes a project rather than a wall. In a market that reorders which skills matter
every few years, the ability to acquire the next one on demand is worth more than
any single skill. Train it deliberately.

One principle governs everything below: stop consuming, start retrieving and
practicing. Rereading, re-watching, and highlighting produce a false sense of
competence. Learning is effortful. You recall information from memory and practice
the skill at the edge of your ability. The sequence below installs that habit in the
fastest order. AI changes how fast you can run the loop, not the loop itself; the
second half of this guide covers that.

```mermaid
flowchart TD
  M([Map the subject]) --> U[Understand]
  U --> R[Retrieve]
  R --> P[Practice at the edge]
  P --> F[Feedback]
  F --> Re[Space and sleep]
  Re -->|next session| R
```

## The path

Follow the order. Each step depends on the one before it.

```mermaid
flowchart TD
  S([Start]) --> A["1. Oakley's talk (10 min)"]
  A --> B["2. Make It Stick + Lobdell"]
  B --> C["3. The Science of Thinking"]
  C --> D["4. Ultralearning + Peak"]
  D --> E["5. The rest, as needed"]
  E --> L([Run the loop])
```

### First: watch [Oakley's talk](/library/oakley-learning-how-to-learn/) (10 minutes)

Watch this first. It removes the belief that some people cannot learn hard subjects.
Barbara Oakley explains the two modes the brain uses to learn: a focused mode for
deliberate work, and a diffuse mode that processes problems during rest. The talk
runs ten minutes and covers the core of *A Mind for Numbers*.

### Second: read [Make It Stick](/library/make-it-stick/)

Read this next. It explains how memory works: retrieval, spacing, interleaving, and
desirable difficulty. Read it actively. At the end of each chapter, close the book
and write down what you remember on a blank page. The rest of this path assumes you
have absorbed it. Read [Marty Lobdell's lecture](/library/lobdell-study-less-study-smart/)
alongside it. It converts the same principles into a daily routine: chunked sessions,
a fixed study location, and recall instead of rereading.

### Third: watch [The Science of Thinking](/library/veritasium-science-of-thinking/)

It explains the two systems of thought: a fast, automatic System 1 and a slow,
effortful System 2. Effortful learning is the only kind that rewires the brain. When
a session feels easy, no learning is happening. Difficulty is the indicator of
progress.

### Fourth: read [Ultralearning](/library/ultralearning/) and [Peak](/library/peak/) together

Ultralearning provides the structure of a self-directed project. Spend the first
tenth of your time mapping the concepts, facts, and procedures of the subject,
identify the 20% of sub-skills that produce most of the result, and learn directly
by doing the real task. Peak defines effective practice: deliberate repetitions past
your current ability, immediate feedback, and isolated drilling of your weakest
sub-skill. Watch [The 4 Things It Takes to Be an Expert](/library/veritasium-4-things-expert/)
with Peak. It makes the same case in eighteen minutes and disproves the
ten-thousand-hours rule.

### Fifth: use the rest as needed

Open [A Mind for Numbers](/library/a-mind-for-numbers/) when a subject intimidates
you. It addresses motivation and tactics. Use [How to Read a Book](/library/how-to-read-a-book/)
as a reference for dense material and apply its four levels of reading. Watch
[Justin Sung's talk](/library/sung-stop-studying-start-learning/) once you have a
working system. It sharpens encoding, so you build durable knowledge structures
rather than relying on flashcards.

## Use AI to run the loop

The principles above predate AI and do not change. What changes is speed. A model
can explain, quiz, simulate, and critique on demand, which collapses the cycle of
explanation, practice, feedback, and review from days into minutes. The danger runs
the other way now: AI makes consuming feel like learning. A tidy summary you nod
along to leaves nothing behind. Point the model at the effortful parts, retrieval
and practice, instead of letting it do them for you. [Co-Intelligence](/library/co-intelligence/)
is the case for treating the model as a working partner.
[A Guide for Thinking Humans](/library/ai-guide-thinking-humans/) is the
corrective that keeps you skeptical of its confident, wrong answers.

### Map the subject (metalearning)

Spend the first tenth of your time drawing the map, now with a research assistant.
Ask for the minimum effective curriculum, the 20% of the subject that unlocks most
of the useful performance. Then have the model split the field into concepts to
understand, facts to memorize, and procedures to practice. Ask how strong university
courses and working experts sequence the same material, then turn that into a dated
plan. Treat every answer as a draft to verify, not a syllabus to trust.

### Build understanding with a Socratic tutor

Expertise is built on mental representations, the internal models you reason
with. Ask the model to explain a concept to a beginner, a practitioner, and an
expert, so you can see where your own understanding stops. Then
run the Feynman technique against it: explain the idea back in your own words and ask
the model to find the logical gaps and the places where you have an illusion of
explanatory depth. Ask for a concrete analogy to anchor anything abstract: voltage
as water pressure, a cache as a desk you keep your most-used files on.

### Drive retrieval and spacing

Passive review is labor in vain. Never end a session without pulling the material
back out of your head. Have the model turn your notes or a source document into ten
hard short-answer questions, then grade your answers. Better, do free recall first:
write everything you remember on a blank page, then ask the model to compare it
against the source and name exactly what you left out. Ground it in your own cited
sources. A tool like NotebookLM keeps the model from inventing facts. Move the
questions into a spaced-repetition system like Anki so they resurface just as you are
about to forget them.

### Practice directly, in simulation

Skill transfers best when you train in the situation you will use it in. Ask the
model to stage the real scenario, whether a negotiation, a diagnosis, a coding
problem, or a hostile code review, and to withhold hints unless you explicitly ask,
so you have to struggle through it. Tell it to play a demanding critic or a high-stakes client
to push you past comfortable. Find the rate-determining step, the one sub-skill
holding the rest back, and have it generate drills that hit only that.

### Get feedback, fast

Immediate, specific correction is the most consistent factor separating experts from
the rest. Submit your work and ask the model to name the three highest-value
mistakes, explain why each one keeps happening, and give you a drill to fix it. Act
on the correction immediately, then resubmit. Stay the human in the loop: you need
enough of your own judgment to catch a confident, wrong answer, which is exactly the
skill *A Guide for Thinking Humans* sharpens. Then step away. Sleep and exercise let
the diffuse mode consolidate the work in the background.

### The loop as a table

Keep this where you can reach it. Each row is one move in the cycle.

| Phase | Principle | Prompt to run |
| --- | --- | --- |
| Map | Metalearning | "Build a 30-day plan for the 20% of [skill] that drives 80% of the result, split into concepts, facts, and procedures." |
| Understand | Elaboration | "Explain [concept] to a beginner, a practitioner, and an expert. Then find the gaps in my explanation below." |
| Retrieve | Active recall | "Turn this into 10 hard short-answer questions, then grade my answers against the source." |
| Practice | Directness | "Simulate [real scenario]. Refuse hints unless I ask." |
| Feedback | Deliberate practice | "Name the 3 highest-value mistakes in my work, why they recur, and a drill for each." |
| Retain | Spacing | "Convert my weak spots into Anki-style spaced-repetition cards." |

## Operating principles

These apply to every subject. Keep them on hand.

- Retrieve before you reread. Free recall outperforms review.
- Space reviews across days and interleave topics. Mixing problem types forces you to identify which method applies.
- Practice at the edge of your ability and seek immediate feedback. Comfortable work produces little. Act on the correction.
- Map before you begin. Spend the first tenth of your time on the concepts, facts, procedures, and the 20% that determines progress.
- Work in focused intervals of twenty-five to fifty minutes on a single task, then rest. Sleep consolidates memory and exercise supports it.
- Teach the material to test it. An explanation a child can follow proves understanding. A gap exposes what to study next.
- Aim AI at the effortful parts, not around them. Make it quiz, simulate, and critique; never let it summarize in place of your own recall. Verify what it tells you.

## The cost

The method is simple. Applying it consistently is not. The principles fit on an index
card. The work is executing them every day for years, while most people continue to
reread and highlight. AI lowers the friction of every step, which makes the trap
sharper, not softer: it is easier than ever to feel productive while learning
nothing. The process is solitary. There is no cohort and no external grading, and
progress slows sharply after the early gains. Improvement requires performing badly
in public while you learn, which most people avoid. The constraint then shifts from
whether you can learn a hard skill to which skill you choose and whether you continue
after the fast progress ends. The books supply the method. Continuing past that point
is the one part they cannot provide.