Andy Matuschak — The reason most learning tools fail

Dwarkesh Patel introduces Andy Matuschak as a researcher, engineer and designer working on tools for thought, and points to a companion video on Andy’s YouTube channel in which Patel watched Andy try to learn new material. The material was an introductory chapter of quantum mechanics. Patel expected to see clever techniques; what surprised him was how deliberate and effortful the practice was. Andy spent about 15 minutes per textbook page, stopping on each small point he did not fully understand, such as whether the relationship the author draws between a classical mechanics equation and a quantum one was really an analogy.

Patel asks whether people in general care about integrating and understanding what they read, and whether they would make the effort if they did. Matuschak says the question is too general to answer directly, but offers two observations. Most students are studying things they do not care about, or they cannot see a clear connection between the reading in front of them and what first inspired them about the subject, so their motivation is always tenuous. Yet when something is truly meaningful to people, their behavior shows they want to absorb it. He points to subreddits where people bring intense, nerdy fascination to gardening equipment or knots, competing to tie some obscure 18th-century knot while flipping through almanacs from the period.

His second observation draws on Adler and Van Doren’s How to Read a Book. Many people keep demanding, aspirational books such as King Lear on the bedside table and fall asleep reading them. The problem is not memory; they are simply not apprehending the words on the page. The authors’ diagnosis, as Matuschak relays it, is that these readers often do not lack the will to stay awake but do not know how to read such a text. He compares them to an inexperienced rock climber facing a wall with only subtle notches: the experienced climber sees a ladder, the novice sees a solid wall. His claim, which he flags as possibly optimistic and open to challenge, is that skillful reading is a real thing, and that when the reading matches a person’s actual interests, greater skill tends in practice to produce more serious, more understanding reading.

Patel frames two models of why people fail to retain what they read: they once understood it and forgot, or they never understood it and never noticed. Watching Andy ask “Have I got this?” sentence by sentence, he found passages from the week before that he had not understood either, without realizing it. Matuschak says noticing confusion is a habit and a skill. Adler and Van Doren’s first rule of active reading is to ask questions and try to answer them, which raises further questions about what to ask and how to answer when the author is not present. He says he read many books the other way, particularly before he built this habit. His problem was not forgetting but that his eyes skidded across paragraphs without his realizing it.

an undemanding reader asks no questions and gets no answers.

Patel widens the question: with online resources like Khan Academy, conscientiousness has become much more valuable, because anyone who can motivate themselves can learn. In the past a professor, peers and in-person deadlines supplied motivation. He asks what design, interface or content changes could give a conscientiousness boost comparable to what pen and paper gives mathematical ability. Matuschak starts from an enduring result in educational psychology: heavy cognition makes metacognition difficult. When you are thinking hard about the page, it is hard to plan, regulate yourself, choose the next action, or evaluate whether you understand, and all of this worsens as the material gets harder and less familiar. One common response is therefore outsourced metacognition. A syllabus that tells you what to read and when is a familiar example, and it is a user-interface and design practice. A self-motivated student can borrow a graduate course syllabus matching the text as a guide to what matters and how to approach it.

Interfaces can build this in directly. In Quantum Country, a textbook on quantum computation that Matuschak developed with Michael Nielsen to explore augmented reading experiences, they embedded review questions roughly every 1500 words. The primary aim was to help people remember. Their theory was that complex subjects are hard partly because many new definitions, notations and terms arrive at once and must be combined while still unfamiliar, so the reader is constantly retrieving these elements, slowly or with a low success rate.

The questions also had a metacognitive benefit. Being asked after 1500 words lets readers notice that they did not absorb the text, not because they forgot but because a word in the question that is evidently important was one they never noticed. That feedback may send them back to reread a section and may change how they read later sections. In interviews, readers said that after the first set of questions, or a particularly difficult set, they slowed down and read more attentively, or realized their reading practices were ineffective in general. Matuschak adds that research on adjunct questions, which accompany a text, shows several effects: the forward-looking effect he describes, and prompting reflection on what was just learned, such as asking why the question matters.

Patel raises a difficulty. A self-learner with a goal may start asking whether a given chapter is really needed, which is exactly the metacognition the syllabus was meant to spare. If a resource closely approximates the syllabus you want, should you follow it without knowing why you need a chapter, or use your own judgment?

Matuschak calls this a classic problem: to bootstrap yourself in a domain you must outsource the question of what is necessary to know. Someone who wants to build an image generator like Midjourney may not know what to study, so they pick up machine learning textbooks and rely on the author’s answer. Even a book labeled as exactly what you need to know still outsources the answer to its author. His advice is to treat that answer as a tentative start, revise it iteratively, and lean on it less as skill grows. He thinks a very common mistake is feeling that one must follow the right, exhaustive, completionist path on faith. When people become bored or unmotivated because the material does not relate to what they want to know, and then stop, they would have been better off skipping around by interest and continuing.

He describes syllabi as scaffolds, a term of art from learning science. Education uses scaffolding everywhere: simpler versions of questions first, partially worked answers, and worked examples where the student predicts the next step. The metaphor breaks down in one respect. As a learner becomes more capable, the scaffolding is removed, a process called fading. After many calculus problems, a half-solved problem with one blank to fill in would no longer be as effective a learning experience.

In computer science, a field he knows well, Matuschak does not need syllabi in the same way. He gives two reasons. The subject places much lower cognitive demand on him, because so much is already familiar that he can plan effectively as he goes. And because he knows so much, he can make a better plan from the start, since planning requires modeling or predicting a path. When the starting point and destination are far apart, the intermediate steps and connections are hard to see; when they are only a couple of hops apart, he can infer them fairly accurately.

Patel infers that someone who wants to learn a subject may do well to take an Intro to X course or textbook, not because it serves their immediate problem but because it supplies the context for the real learning. Matuschak agrees and adds that learners do not even know everything that exists, which is another reason we outsource these choices. He sees a basic tension in unschooling. Letting children pursue their interests has real merits, but a child whose true passion turns out to be ocean geology, in a landlocked country where nobody talks about it, misses a great opportunity. A school that brings in a guest speaker for a special lecture on ocean geology could light up that child’s world, even though the child would never have chosen the lecture. Patel says unschooling is a subject he wants to return to.

Patel quotes a passage from Paul Graham’s essay How You Know and asks whether it is acceptable that we forget so much of what we read.

Your mind is like a compiled program you’ve lost the source of. It works, but you don’t know why.

Matuschak says the claim is true to some extent, and whether that extent is enough depends on the situation. If your aspiration depends on a deep, detailed understanding of the material, the imprint a book leaves on your worldview or automatic responses may not be sufficient. If what you want is to absorb many ways of looking at the world, the details matter less. Knowing that Confucius emphasizes community and society as a moral patient, in contrast to the individualism of a group of humanist philosophers, may be all you need if that is the level at which you make decisions in that domain.

He notes that the word “compile” echoes a prominent theory of cognition, John Anderson’s ACT-R, whose process of knowledge compilation turns individual facts into higher-level patterns that can be generalized and applied in more contexts. Reading a story or case study teaches some generalization this way. Anderson’s theory makes specific claims about what compilation requires, and Matuschak expects Anderson would say that turning learned material into schemas that match future scenarios needs repeated exposure, actual use, and a variety of activities that show the brain the material is relevant in combination. Simply reading probably will not do that. Reading combined with many conversations, in a context that demands and draws on what you read, might produce the compilation step naturally.