I've written in the past a couple of blogposts on education and learning: these two on Bloom's two sigma and mastery learning, and spaced repetition systems (SRS) are the main examples.

Something I noticed today is that on the one hand I have these blogposts, and I find SRS valuable in theory but in practice what I do is something quite different: massive input. I refer to this concept in passing in Scaling tacit knowledge

The interesting thing of language learning is how effortless it seems to be for children. The conjunction of massive input of examples with the right context leads initially to remember salient words first, then noticing overall patterns, inferring grammar, and ultimately speaking the language proficiently. Adults can learn languages in the same way in about a year by the same means: exposure to a large library of examples with the right context. In one case, 18 months was enough for this one person to go from zero to near-native proficiency in Japanese.

I am not claiming we can learn everything using the same mental structures we use for language. Perhaps language is easier than other domains because we are pre-wired for language acquisition in a way we are not for other domains. I am saying that there is a domain where this (massive input of examples with context) obviously works and we should think about seeing if we can expand that to other domains.

Massive input is exactly what I did when I wrote the Longevity FAQ: I read a textbook, then read a lot of papers over many months, not stopping to read them particularly carefully or taking any notes, I tried to maximize quantity over depth for any one specific paper. And as I mention in the quote above, massive input as a concept is coming from the world of language acquisition, the place where most people are first exposed to flashcards (or some basic form of spaced repetition), which I find quite interesting.

This is the extreme opposite from the way Andy Matuschak describes reading a book on quantum mechanics. To quote from there:

Dwarkesh was quite surprised by my approach to the book. I moved at a pace of about fifteen minutes per page, while he had spent a few minutes or less. More importantly, I was constantly asking questions of the text and of myself. Some examples:

  • What does this sentence mean? Can I explain it in my own words?
  • Which ideas are particularly important here?
  • The author clearly thinks I should see why this claim is true—so why is it true?
  • The author’s emphasizing this detail—so why is it important?
  • The author seems to be setting up a contrast here—so what is it, exactly?
  • How does this detail relate to my prior knowledge in physics?
  • If I hide all but the beginning of this worked example, can I produce the rest myself?
  • I made a mistake a moment ago—do I understand why? Can I explain my misapprehension?
  • And of course: can I simply recall what was said on the previous page?

The advantage of massive input is that you don't need to force yourself to stop to ask all these questions (ie forcing yourself to slow down) or to slow down to write notes. In my own experience learning biology, the reason why this worked for me is that when starting in a new field, there are a lot of things that are unclear to the novice if they are important, even things in textbook. Maybe nothing else you'll ever see again will leverage that concept. That's highly likely in bio. When you know that most of what you are reading may not be important, forcing yourself to take notes and make SRS prompts makes the learning process slower and not rewarding at all. The massive input approach gives you permission to skim, jump ahead, and go on tangents.

But in contrast, I would never do this with any formal subject (math, physics, learning a new programming language, computer science in general). There I think Andy's approach is the right one and the one I've tried to follow in the past.

The why seems clear: Formal domains tend to rely a lot on deep towers of abstraction, mastering one level makes a lot of sense to get to the next one because the next one is the previous one, wrapped in new symbols. The number of concepts encountered is small, but they way they can interact is very rich, and one has to be able to work through these interactions as part of producing output in the domain in question. Whereas other domains like biology have less abstraction. The underlying nature of the domain is extremely complex, the nature of the interactions of the components in biological systems are not well defined by the sort of formal rules one can neatly encapsulate in abstractions, it all leaks to some extent; so in practice in biology one is not so much proving things but gesturing in various directions and then pinning down the gesturing with concrete experiments.

And then there's language learning.

"Khatzumoto", the author of AJATT, or "All Japanese All The Time" is one such proponent of massive input. It worked out for him, here you can see him speaking really good Japanese

I learned Japanese in 18 months. In June 2004, at the ripe old age of 21, all post-pubescent and supposedly past my mental prime, I started learning Japanese. By September 2005, I had learned enough to read technical material, conduct business correspondence and job interviews in Japanese. By the next month, I landed a job as a software engineer at a large Japanese company in Tokyo (yay!).

How did I do it? Well, by spending 18-24 hours a day doing something, anything in Japanese ("all Japanese, all the time"). That sounds like a lot of time to invest, but I was almost as busy as you are: a full-time student majoring in computer science at a university in a small town in the US, physically far from Japan and Japanese people. I had computer science coursework, jobs and even a non-Japanese "significant other". In other words, I had a life.

Perhaps this case is a bit extreme, but you get the point. Massive input can be combined with SRS, but part of the point is that you don't have to. If all your day revolves around the topic, then concept seen a week ago is likely to come back this week, so you are effectively SRSing yourself without intending to, without making prompts. Indeed in massive input land you can find opposition to SRS as getting in the way of massive input! One can of course also find support for a moderate position where one does a bit of both.

Effectiveness aside, one recurrent element through Andy's thinking on designing ways to make what we read stick is the coerciveness of SRS: you have to force yourself to stick to the system. Can that coerciveness be reduced and made gentle? A nice narrative for why one is doing SRS is a first step: "If I do X then I'll get benefits Y". If you observe other people that are having success (getting Y) with SRS then you will have motivation to keep grinding at it: they did it and so can you. But in most domains one doesn't have such reassurances. Again: if you tried to Anki the entirety of The Molecular Biology of the Cell, do you think you'll be more effective than if you read through it and then read a couple of papers?

So the recommended approach: Build a map of the domain with massive input, immerse yourself in the domain. Only then later apply spaced repetition for those final missing bits.