Thoughts on Conceptual Learning, Anki, Books, and Physicality

Conceptual Learning Issues

I want to spell out what I see as some core issues with learning abstract, ladder-like concepts—think of mathematics (algebra to calculus) or physics building up a more complex picture from simpler parts.

When learning a new topic, whether from a textbook, article, or podcast, there are always aspects of the experience that have, for me, been a struggle. When wading into a new field, you face an onslaught of terminology and nested abstractions. This concept relies on that other concept, which sits on some earlier idea—all labelled with unique vocabulary.

Additionally, I’m an avid user of Anki, the open-source spaced-repetition application. I have approximately 20,000 cards, 95% created by myself.

Naturally, to ease the pain of learning a new conceptual field—or at least mitigate some of the hardest aspects—I’ve attempted to use spaced repetition to catalogue my journey and remind me of the core concepts. But outside of vocabulary and grammar for languages, I’ve found it incredibly difficult to maintain effective usage of spaced repetition for conceptual or contingent knowledge. Ideas that depend on other ideas, or formulas that build on earlier intuition.

Why should this be?

Why Should This Be?

It has puzzled me. I get great value from spaced repetition in one area—languages—but outside of that it fails completely. I know it’s not a memory issue. With 20,000 cards, I have remarkably few retention problems.

For language cards there is a lot of reverse data: learning a word in one direction and in the other direction. So if you want to count unique, non-reverse cards, there are roughly 10,000.

So what do I find so difficult about abstract-ladder (AL) knowledge when placed into the spaced-repetition format of Anki?

For one, the spatial relationship between pieces of knowledge is lost once placed in the context of a single, lone card floating in the ether. True, it’s possible to organize cards, but depending on your recall, responses, or deck-selection habits, the cards themselves will never have consistent spatial relations.

There are card-specific adjustments like cloze deletions, where you hide some aspect of an image or sentence to then reveal it on review. But these create spatial relations within a single card—not between pieces of knowledge.

What Would a Solution Look Like?

If we look outside the world of spaced repetition, there are hints at possible solutions. In the humble form of a physical book, you have natural spatial organization of knowledge. And to a surprising degree, when discussing how people learn technical topics, a consistent theme is the use of physical books over digital ones. The digital book is useful once the concept is understood or the general field is known. But the initial hard ingestion of information seems easier in a physical form. A well-constructed book also exhibits a form of spaced repetition, with later sections referring back to earlier concepts.

There are also examples of knowledge progression and layered abstractions in gaming. Even in simple games, you have a gameplay loop where you learn a technique or concept and are then asked to use it to solve a problem. Over the life of the game, new ideas, mechanics, and concepts are introduced, leading to a surprisingly dense buildup of abstraction from the initial ideas.

Whether playing Minecraft, a city-builder, or a platformer with combo moves, there is a remarkable degree of complexity and abstraction conveyed through this spatially arranged loop of progression, challenge, and reminder.

Another area where learning works well is in context switching. When reading, you typically want to continue reading. It feels good; there’s progression; you feel like you’re achieving something. But for learning concepts, you need to be questioned and challenged to use them. Unfortunately, the emotional pull to just keep reading often wins, even though I know exercises are useful. My mind says: it’s boring, it’s annoying, leave me alone and let me read.

But this resistance can be short-circuited—or tricked—by switching context entirely.

Take the example of reading a book on electronics. It describes the functioning of a circuit and some associated programming. You then switch contexts completely to physically building the mini-circuit and programming it. Why does this feel less burdensome and less likely to trigger resistance? Functionally, you’re doing the exercise—implementing the concept.

I think the reason this works is that the context has shifted enough that your mind treats the two activities as independent streams of progression, with some casual cross-linkage. You’re not doing exercises that feel like a test of your reading comprehension. Instead, you’re creating something new that can’t be judged by the same metrics.

Conclusion (For Now)

I think there’s a possible way forward—a landscape of solutions—where you can retain the benefits of spaced-repetition tooling and the spatially related knowledge found in physical books, while vastly improving over both. To hint at it: I think the solution lies in using the physical intuition of your own mind, the flexibility of computing platforms, and the interlinking capabilities of large language models.