I am a slow reader compared to most people I know, and have occasionally looked into speed reading to try to remedy this.
I was surprised about your description of speed reading as “eye technique”, because although I’ve seen this sort of thing, I’ve also seen mental processing advice. If someone asked me to quickly summarize what I know about speed reading, I would have said “Speed reading guides usually say that you have to stop turning the words into imaginary auditory signals in a mental voice; this forces you to slow down, as you first convert the letters to mental sounds, and then convert the sounds to meaning. Instead, you should interpret the letters directly into words and meaning, without involving the auditory cortex.”
(I’ve never succeeded in doing this, but it’s led me to practice EG repeating nonsense words in my head while I read, to try and force myself not to use auditory processing.)
Anyway.
I think it’s important to point out the difference between optimal comprehension per character vs optimal comprehension per second. A natural hypothesis to me (alternate to your natural pace hypothesis) is that people intuitively optimize comprehension per character, because it’s more salient (they’re not watching the clock as they read, after all).
On this hypothesis, most people will be a bit too low on the comprehension curve, because they aren’t automatically optimizing for that.
Either way, I find it a priori very plausible that people aren’t at max reading skill automatically, and can benefit from focused practice. Although I haven’t noticed marked improvements, myself. I definitely endorse your re-framing in terms of optimizing comprehension rate rather than raw reading speed.
One of the most subjectively helpful frames (which I got from someone at a LessWrong meetup) was that speed reading is about optimizing information absorbtion. Notice when you already get what the author is trying to say. If you have a guess about what argument the next paragraph is going to make, spot-check your guess by reading a few phrases; if they’re consistent with your expectations, you can probably skip ahead. Lots of people sort of “cargo cult” reading books, by reading the whole thing in order. It’s often more efficient to skip around and find the information you actually want.
I think part of what’s going on is that speed-reading advocates know this, and are really advocating techniques which encourage people to skip over things. For example, you mention:
Regressions: how often the eye travels to an earlier point in the text.
Avoiding regressions means you avoid re-reading stuff you just passed over. I re-read all the time, when I didn’t really get what was said. Someone who advocates avoiding regressions is implicitly telling me: “Oh, you know what they said. You’re probably re-reading just to check. Even if you misunderstood something, you’ll probably be able to figure out what’s really going on based on the next paragraph.”
Which is not necessarily true. Or, even if true, is not necessarily optimal information absorbtion (because the risk of misunderstanding is pretty high with this sort of skimming).
In other words, I reject the idea that “eye technique” is really what it purports to be—better general eye technique for reading. I think it’s actually implicit advice to skim more and fill in details by inference, cashed out as details about eye movements which sorta force you to do this.
I take it as a big compliment that you wrote such a long and thoughtful reply to my post! Thank you!
The distinction you draw between broad vs. deep readers is the reason I didn’t operationalize comprehension. “Comprehension” is just the type of understanding you want to extract from the text on your particular read-through, defined as you please. Maybe let’s think of them in terms of abstract units called Comprehendalons, analogous to a Utilon.
You could define a Deep-Comprehendalon as knowing “wrong with the abstract, and what they think the real conclusions are,” in which case a very slow reading speed is ideal. An ideal reading speed for Deep-Comprehendalons might be 10 wpm, or even slower, and it might take you several hours to acquire just one.
A Shallow-Comprehendalon might be picking up an single atomic fact. An ideal reading speed for Shallow-Comprehendalons might be 500 wpm, or even faster, and you might be able to pick up a huge number of them in a short period of time.
One thing I infer from this framework is that Shallow-Comprehendalons don’t add up to Deep-Comprehendalons. They are not fungible, not the same type of good. Optimizing for one may mean sacrificing the other.
However, that seems debatable. Even if it’s true, the Comprehension Curve would still hold. You’d just have a different ideal reading speed and maximum comprehension rate for each type.
Interestingly, for me, fully engaging my auditory cortex has, I believe, really helped me to move closer to my maximum comprehension rate. I’ll describe this in a future post. One of my motivations for writing this is that I think that speed reading advocates are doing something fundamentally good—experimenting—but doing it in a screwy way, where they invent a whole theory of why their method is the best, without exploring contrasting hypotheses. And then when sober scientists get to studying it, they approach the field by testing the claims of speed readers, rather than by reflecting a priori on what approach to learning ought to enhance comprehension. The latter is what I’ve tried to advance toward here.
I appreciate you bringing up the point that much of speed reading advice revolves not around eye technique, but around mental technique—the bypassing of the auditory cortex. That’s true, and I just entirely left that out for no good reason. I’m going to edit the OP to include a reference to it, along with a credit to you for reminding me of it.
One of my future posts will discuss what I’ve noticed in regards to skimming. I’ll sum it up for you, as practice and since you brought it up.
Let’s consider the following sentence from a biochemistry textbook:
“Oxidative reactions of the citric acid cycle generate reduced electron carriers that are then reoxidized to drive the synthesis of ATP.”
For someone like myself who’s familiar with biochemistry, many of the individual words and phrases refer to concepts that I already understand well. But there are particular keywords within the sentence that “focalize” a new concept, build out of the others. Let me break down my experience reading it:
“Oxidative reactions”—the word oxidative is key, and “contains” the concept of a reaction within it.
“of the citric acid cycle”—citric is key, and automatically refers to “citric acid cycle”
“generate reduced electron carriers”—reduced is key, and “generate” is implied by the connection between “citric [acid cycle]” and “reduced”
“that are then reoxidized”—reoxidized is key
“to drive the synthesis of ATP.”—ATP is key, because I already know that the end result is to synthesize, rather than consume ATP.
So as I read this sentence, the words “oxidative,” “citric,” “reduced,” and “ATP” get lodged in my working memory, repeated in my auditory cortex in a sort of earworm-like jingle. While they repeat, my eyes scan the rest of the words to observe how they link up. So I read with a two-layered awareness: the working memory jingle-words that isolate and relate the key concepts, and the non-auditory connecting words that give them meaning and relate them together. This is just the approach to reading that I’m playing with right now, and I make no claim that it’s useful or ideal for myself or anybody else.
But it’s an interesting riff on skimming and speed reading. Instead of divorcing myself from my auditory cortex, I use it for keywords only, while relying on non-auditory reading for connecting words, which I can largely skim through.
The only way to develop ideas like this is to experiment openly, with a goal not of reading quickly, but of being experimental with your approach and trying to intuitively feel your way toward a method that is satisfying and feels like you’ve comprehended the material well. I find that this approach makes it far easier to pay nuanced attention to the material, read for long stretches of time without fatigue, and relate concepts.
Imagine for the sake of argument that we’re happy to just absorb information, with no topic prioritization; we just want our map to get closer to the territory.
There’s still the question of exactly what our loss function is: how much do we like being a specific distance from the truth? Our loss function could be more like Bayes loss, which punishes overconfidence highly (assigning probability close to zero for an event which actually is true gets us an arbitrarily bad score); or, it could be more like Brier score (which caps the amount you can lose for any particular wrong idea).
I think deep readers are more like Bayes-score maximizers: skimming the abstract and getting wrong information feels like a big risk. No quantity of improved beliefs can necessarily compensate for one mistake, because mistakes can be arbitrarily bad. A bayes-score maximizer skimming something is conscious of exactly how many grains of salt go with each thing learned, because getting that wrong can be very costly—so they would feel a need to remember “I was speed-reading this, so I should doubt all my conclusions about it a little bit more”.
Broad readers are more like Brier-score maximizers: skimming the abstract and getting a wrong conclusion is only a bounded risk, so it’s easily balanced by the benefit of lots of knowledge. They don’t feel an overwhelming need to count grains of salt, because 95% wrong is not that different from 100% wrong; so they happily accept a bunch of noisy information in, without worrying too much about careful tabulation of the noise.
I don’t have too many intuitions about when Bayes score or Brier score will be closer to our true utility-of-knowledge functions. But I suspect deep reading is more useful for the kind of research where you’re trying to generate really new things, like totally new hypotheses or new areas of mathematics. Whereas broad reading may be more useful for “applied” type research, where you’re taking existing knowledge and using it in new areas.
If we assume that the accuracy improvements to researching a given question are logarithmic, then it would make sense to read broadly on unimportant questions and read deeply on crucial questions.
Signaling also seems relevant here. It might be advantageous to be widely informed, or to be seen as the kind of person who only speaks on their domain of expertise.
There could also be times when you just need to be conversant in the subject enough to know who to delegate the deeper research to.
So in general, I would expect the value of broad vs. deep research to be highly contextual.
But I wonder if the same habits that may lead people to anchor on an inappropriate reading speed also lead them to anchor on a sometimes inappropriate reading depth. It’s plausible to me that people who tend to read broadly by habit could reap significant gains by practicing deep reading on even an arbitrary subject, and vice versa.
It would be interesting if there was an equivalent to the DSM, but for reading habits. Could we imagine a test or a set of diagnostic criterion that could classify people both according to their level of reading proficiency, and also according to their habitual level of depth/breadth? So for example, a low-skill but deep reader might be a religious fundamentalist who has their text of choice practically memorized, yet who has very little familiarity with the nuances of interpretation. By contrast, we can imagine low-skill broad readers, who read all kinds of novels and newspapers but remember very little of it. And high-skill broad or deep readers, of course.
I think this is related to one of my perennial topics of interest, which is the path toward a specialization. A science student in undergrad or earlier reads broadly about science. At some point, if they continue on a scientific path, they eventually focus on a much narrower area, and their whole reading program focuses on acquiring knowledge that they perceive as directly useful to a specific research project.
As I’ve graduated into this phase, I’ve found that the deep, related, specialized, purposeful reading is vastly more satisfying than the broad, shallow, disconnected reading that came before. It makes me suspect that one reason people get turned off of science early is that they never get the experience of “cocooning” in a specialty in which all the articles you read are riffing off each other, interrelated, and building toward a goal. It’s the closest thing that I’ve found to programming, which also entails building an interrelated construct to make predictions and do useful work.
I’m also interested in whether and how “broad reading” can be done with an equivalent sense of purpose. There’s an article on Applied Divinity Studies, Beware the Casual Polymath, which I think can be characterized as a criticism of superficially high-skilled, but in fact low-skilled broad readers. It’s pointing out that just because you’re reading all kinds of Smart People Stuff doesn’t mean that you’re actually learning effectively.
I would imagine that a high-skilled broad reader would be somebody who has a role that involves lots of delegation and decision-making. The fictional example that comes to mind is a member of the Bartlett senior staff in the West Wing, who have to understand a huge number of issues of national significance, but only just enough to know who to delegate to or what positions are at least not-insane. For them, making reasonable, if not necessarily perfectly optimized, choices, but making a decision, is much more important than getting the exact right answer. So I would describe them as a depiction of a high-skilled, broad reader.
I am a slow reader compared to most people I know, and have occasionally looked into speed reading to try to remedy this.
I was surprised about your description of speed reading as “eye technique”, because although I’ve seen this sort of thing, I’ve also seen mental processing advice. If someone asked me to quickly summarize what I know about speed reading, I would have said “Speed reading guides usually say that you have to stop turning the words into imaginary auditory signals in a mental voice; this forces you to slow down, as you first convert the letters to mental sounds, and then convert the sounds to meaning. Instead, you should interpret the letters directly into words and meaning, without involving the auditory cortex.”
(I’ve never succeeded in doing this, but it’s led me to practice EG repeating nonsense words in my head while I read, to try and force myself not to use auditory processing.)
Anyway.
I think it’s important to point out the difference between optimal comprehension per character vs optimal comprehension per second. A natural hypothesis to me (alternate to your natural pace hypothesis) is that people intuitively optimize comprehension per character, because it’s more salient (they’re not watching the clock as they read, after all).
On this hypothesis, most people will be a bit too low on the comprehension curve, because they aren’t automatically optimizing for that.
Either way, I find it a priori very plausible that people aren’t at max reading skill automatically, and can benefit from focused practice. Although I haven’t noticed marked improvements, myself. I definitely endorse your re-framing in terms of optimizing comprehension rate rather than raw reading speed.
One of the most subjectively helpful frames (which I got from someone at a LessWrong meetup) was that speed reading is about optimizing information absorbtion. Notice when you already get what the author is trying to say. If you have a guess about what argument the next paragraph is going to make, spot-check your guess by reading a few phrases; if they’re consistent with your expectations, you can probably skip ahead. Lots of people sort of “cargo cult” reading books, by reading the whole thing in order. It’s often more efficient to skip around and find the information you actually want.
I think part of what’s going on is that speed-reading advocates know this, and are really advocating techniques which encourage people to skip over things. For example, you mention:
Avoiding regressions means you avoid re-reading stuff you just passed over. I re-read all the time, when I didn’t really get what was said. Someone who advocates avoiding regressions is implicitly telling me: “Oh, you know what they said. You’re probably re-reading just to check. Even if you misunderstood something, you’ll probably be able to figure out what’s really going on based on the next paragraph.”
Which is not necessarily true. Or, even if true, is not necessarily optimal information absorbtion (because the risk of misunderstanding is pretty high with this sort of skimming).
In other words, I reject the idea that “eye technique” is really what it purports to be—better general eye technique for reading. I think it’s actually implicit advice to skim more and fill in details by inference, cashed out as details about eye movements which sorta force you to do this.
I take it as a big compliment that you wrote such a long and thoughtful reply to my post! Thank you!
The distinction you draw between broad vs. deep readers is the reason I didn’t operationalize comprehension. “Comprehension” is just the type of understanding you want to extract from the text on your particular read-through, defined as you please. Maybe let’s think of them in terms of abstract units called Comprehendalons, analogous to a Utilon.
You could define a Deep-Comprehendalon as knowing “wrong with the abstract, and what they think the real conclusions are,” in which case a very slow reading speed is ideal. An ideal reading speed for Deep-Comprehendalons might be 10 wpm, or even slower, and it might take you several hours to acquire just one.
A Shallow-Comprehendalon might be picking up an single atomic fact. An ideal reading speed for Shallow-Comprehendalons might be 500 wpm, or even faster, and you might be able to pick up a huge number of them in a short period of time.
One thing I infer from this framework is that Shallow-Comprehendalons don’t add up to Deep-Comprehendalons. They are not fungible, not the same type of good. Optimizing for one may mean sacrificing the other.
However, that seems debatable. Even if it’s true, the Comprehension Curve would still hold. You’d just have a different ideal reading speed and maximum comprehension rate for each type.
Interestingly, for me, fully engaging my auditory cortex has, I believe, really helped me to move closer to my maximum comprehension rate. I’ll describe this in a future post. One of my motivations for writing this is that I think that speed reading advocates are doing something fundamentally good—experimenting—but doing it in a screwy way, where they invent a whole theory of why their method is the best, without exploring contrasting hypotheses. And then when sober scientists get to studying it, they approach the field by testing the claims of speed readers, rather than by reflecting a priori on what approach to learning ought to enhance comprehension. The latter is what I’ve tried to advance toward here.
I appreciate you bringing up the point that much of speed reading advice revolves not around eye technique, but around mental technique—the bypassing of the auditory cortex. That’s true, and I just entirely left that out for no good reason. I’m going to edit the OP to include a reference to it, along with a credit to you for reminding me of it.
One of my future posts will discuss what I’ve noticed in regards to skimming. I’ll sum it up for you, as practice and since you brought it up.
Let’s consider the following sentence from a biochemistry textbook:
“Oxidative reactions of the citric acid cycle generate reduced electron carriers that are then reoxidized to drive the synthesis of ATP.”
For someone like myself who’s familiar with biochemistry, many of the individual words and phrases refer to concepts that I already understand well. But there are particular keywords within the sentence that “focalize” a new concept, build out of the others. Let me break down my experience reading it:
“Oxidative reactions”—the word oxidative is key, and “contains” the concept of a reaction within it.
“of the citric acid cycle”—citric is key, and automatically refers to “citric acid cycle”
“generate reduced electron carriers”—reduced is key, and “generate” is implied by the connection between “citric [acid cycle]” and “reduced”
“that are then reoxidized”—reoxidized is key
“to drive the synthesis of ATP.”—ATP is key, because I already know that the end result is to synthesize, rather than consume ATP.
So as I read this sentence, the words “oxidative,” “citric,” “reduced,” and “ATP” get lodged in my working memory, repeated in my auditory cortex in a sort of earworm-like jingle. While they repeat, my eyes scan the rest of the words to observe how they link up. So I read with a two-layered awareness: the working memory jingle-words that isolate and relate the key concepts, and the non-auditory connecting words that give them meaning and relate them together. This is just the approach to reading that I’m playing with right now, and I make no claim that it’s useful or ideal for myself or anybody else.
But it’s an interesting riff on skimming and speed reading. Instead of divorcing myself from my auditory cortex, I use it for keywords only, while relying on non-auditory reading for connecting words, which I can largely skim through.
The only way to develop ideas like this is to experiment openly, with a goal not of reading quickly, but of being experimental with your approach and trying to intuitively feel your way toward a method that is satisfying and feels like you’ve comprehended the material well. I find that this approach makes it far easier to pay nuanced attention to the material, read for long stretches of time without fatigue, and relate concepts.
Imagine for the sake of argument that we’re happy to just absorb information, with no topic prioritization; we just want our map to get closer to the territory.
There’s still the question of exactly what our loss function is: how much do we like being a specific distance from the truth? Our loss function could be more like Bayes loss, which punishes overconfidence highly (assigning probability close to zero for an event which actually is true gets us an arbitrarily bad score); or, it could be more like Brier score (which caps the amount you can lose for any particular wrong idea).
I think deep readers are more like Bayes-score maximizers: skimming the abstract and getting wrong information feels like a big risk. No quantity of improved beliefs can necessarily compensate for one mistake, because mistakes can be arbitrarily bad. A bayes-score maximizer skimming something is conscious of exactly how many grains of salt go with each thing learned, because getting that wrong can be very costly—so they would feel a need to remember “I was speed-reading this, so I should doubt all my conclusions about it a little bit more”.
Broad readers are more like Brier-score maximizers: skimming the abstract and getting a wrong conclusion is only a bounded risk, so it’s easily balanced by the benefit of lots of knowledge. They don’t feel an overwhelming need to count grains of salt, because 95% wrong is not that different from 100% wrong; so they happily accept a bunch of noisy information in, without worrying too much about careful tabulation of the noise.
I don’t have too many intuitions about when Bayes score or Brier score will be closer to our true utility-of-knowledge functions. But I suspect deep reading is more useful for the kind of research where you’re trying to generate really new things, like totally new hypotheses or new areas of mathematics. Whereas broad reading may be more useful for “applied” type research, where you’re taking existing knowledge and using it in new areas.
If we assume that the accuracy improvements to researching a given question are logarithmic, then it would make sense to read broadly on unimportant questions and read deeply on crucial questions.
Signaling also seems relevant here. It might be advantageous to be widely informed, or to be seen as the kind of person who only speaks on their domain of expertise.
There could also be times when you just need to be conversant in the subject enough to know who to delegate the deeper research to.
So in general, I would expect the value of broad vs. deep research to be highly contextual.
But I wonder if the same habits that may lead people to anchor on an inappropriate reading speed also lead them to anchor on a sometimes inappropriate reading depth. It’s plausible to me that people who tend to read broadly by habit could reap significant gains by practicing deep reading on even an arbitrary subject, and vice versa.
It would be interesting if there was an equivalent to the DSM, but for reading habits. Could we imagine a test or a set of diagnostic criterion that could classify people both according to their level of reading proficiency, and also according to their habitual level of depth/breadth? So for example, a low-skill but deep reader might be a religious fundamentalist who has their text of choice practically memorized, yet who has very little familiarity with the nuances of interpretation. By contrast, we can imagine low-skill broad readers, who read all kinds of novels and newspapers but remember very little of it. And high-skill broad or deep readers, of course.
I think this is related to one of my perennial topics of interest, which is the path toward a specialization. A science student in undergrad or earlier reads broadly about science. At some point, if they continue on a scientific path, they eventually focus on a much narrower area, and their whole reading program focuses on acquiring knowledge that they perceive as directly useful to a specific research project.
As I’ve graduated into this phase, I’ve found that the deep, related, specialized, purposeful reading is vastly more satisfying than the broad, shallow, disconnected reading that came before. It makes me suspect that one reason people get turned off of science early is that they never get the experience of “cocooning” in a specialty in which all the articles you read are riffing off each other, interrelated, and building toward a goal. It’s the closest thing that I’ve found to programming, which also entails building an interrelated construct to make predictions and do useful work.
I’m also interested in whether and how “broad reading” can be done with an equivalent sense of purpose. There’s an article on Applied Divinity Studies, Beware the Casual Polymath, which I think can be characterized as a criticism of superficially high-skilled, but in fact low-skilled broad readers. It’s pointing out that just because you’re reading all kinds of Smart People Stuff doesn’t mean that you’re actually learning effectively.
I would imagine that a high-skilled broad reader would be somebody who has a role that involves lots of delegation and decision-making. The fictional example that comes to mind is a member of the Bartlett senior staff in the West Wing, who have to understand a huge number of issues of national significance, but only just enough to know who to delegate to or what positions are at least not-insane. For them, making reasonable, if not necessarily perfectly optimized, choices, but making a decision, is much more important than getting the exact right answer. So I would describe them as a depiction of a high-skilled, broad reader.