I think rationalists systematically overestimate the benefits of public discussion (and underestimate the costs).
When something is communicated one-on-one, the cost is effectively doubled because it consumes the sender’s time. I think is roughly offset by a factor of 2 efficiency bonus for tailored/interactive talk relative to static public talk. Even aside from the efficiency of tailored talk, a factor of 2 isn’t very big compared to the other costs and benefits.
So I think the advantages of public discussion are not mostly the efficiency from many people being able to read something that got written once. Instead they are things like:
One-on-one explanations often require the listener to compensate the explainer for their time. Right now we have no good institutions for this problem, so it relies on the explainer caring about what the listener believes or expecting to receive something in return.
By investing in really great explanations you can save total time compared to one-on-one explanations. I think a very small fraction of public discussion falls into this category / meets this bar.
Relative to the potential of both modes, we are currently worse (!) at one-on-one transmission than written broadcasting. For example, we have larger unnecessary overhead and political diversion from one-on-one transmission.
Broadcasting is good for transferring information from people with more valuable time to people with less valuable time—in those cases “using up the explainer’s time” much more than doubles the cost. I expect it would be better to do transmitting information 1-on-1 transmission to specialized explainers who broadcast it, but we aren’t good at that.
Often one-on-one discussion creates artifacts that are cheap to share, and at that point you might as well share (even though it’s a minority of the value and so is only worth doing if it doesn’t increase the cost much). For example, it’s extremely valuable to sort through the evidence and clarify our thinking, and once we do that we its relatively cheap to share. And discussions often produce transcripts that are worth sharing (though there is often political overhead from publishing a discussion, which often outweighs the benefits).
Public discussion serves important non-information-transmission functions, e.g. building consensus, political maneuvering, norm enforcement, etc. In practice I think these are more important than information transmission.
I do think that it’s great to iterate on and improve the rationalist crowd’s fora for public discussion (and I think use of FB is evidence for an unmet need). But:
It’s easy to overestimate the intrinsic benefits of public discussion and so be overly surprised when it doesn’t catch on. (Or to make overly strong inference about groups given the lack of useful public discussion.)
I expect there would be larger returns to improving our local institutions for one-on-one information transmission.
Given that rationalists estimate the value of public discourse (and also private discourse) higher than almost anyone else, it is certainly plausible that we overestimate it. Given I estimate it higher than most rationalists, it’s even more plausible that I overestimate it. But if anything, I think we underestimate it, and Paul points to some of the reasons why.
Paul is implicitly, I believe, employing a sort of ‘student-teacher’ model of communication; the teacher transmits communication, the student learns it. The customization and lack of politicization in person is about a factor of two better, which then roughly ‘cancels out’ the increased cost. But if that’s true, it’s then pointed out that there are a number of other key advantages to public discourse. Paul’s extra reasons seem super important, especially if we add creation of common knowledge and vocabulary to the last one. He also points out that teacher time is typically more valuable than student time, meaning the cost ratio is more than two, perhaps much more. And time spent writing dwarfs the time each person spends reading, so that further increases the ratio.
So it seems like even if Paul’s estimates are complete and right, we should still default to lots of public discourse?
To that I’d add at least:
Reference (we can cite the past as evidence slash explanation that allows us to build upon it, which is super valuable even when almost no one clicks on the link)
Error correction and feedback (if we do 1-on-1, and we get it wrong, we don’t get to fix mistakes or get help from others), review (we can go back and remind ourselves)
Selection (the best explanations and teachers get used more)
Opportunity (if something is being taught in private, I often won’t even know it’s a thing one could learn, at all)
Learning through writing (I learn a ton from having to write down my ideas robustly)
Better crafted explanations by taking the time to get them right and general
Asynchronous and on demand (I can teach today when I have time, or over two weeks as thoughts come to me, then people can learn when they need to learn and have time, which radically decreases effective cost) which only gets more important as people comment and converse slowly over time.
+1 on “learning through writing,” this factor alone is my main disagreement with Paul’s comment. For most people the biggest low-hanging improvement to their thinking might just be to write it down. Once it’s written down it’s not nearly that much effort to edit enough to post in a public discussion.
Zvi already mentioned this, but I just want to emphasize that I think error finding and correction (including noticing possible ambiguities and misunderstandings) is one of the most important functions of discussion, and it works much better in a group discussion (especially for hard to detect errors) because everyone interested in the discussion can see the error as soon as one person notices it and points it out. If you do a series of one-on-one discussions, to achieve the same effect you’d have to track down all the previous people you talked to and inform them. I think in practice very few people would actually bother to do that, due to subconscious status concerns among other reasons, which would lead to diverging epistemic states among people thinking about the topic and ensuing general confusion.(Sometimes an error is found that the “teacher” doesn’t recognize as an error due to various cognitive biases, which would make it impossible to track down all the previous discussants and inform them of the error.)
Zvi’s “asynchronous and on demand” is also a hugely important consideration for me. In an one-on-one discussion, the fact that I can ask questions and get immediate responses is often counterbalanced by the feeling that I don’t have time to think, digest an idea, and decide whether it makes sense or not, because the other person is waiting for me to nod so they can move on (and that person’s time is usually pretty valuable). So frequently I just nod and think that I’ll figure it out later on my own time, but that means I’m not actually taking advantage of the main benefit of having a one-on-one discussion. I think Paul might not be taking this into account when he says “a factor of 2 efficiency bonus for tailored/interactive talk relative to static public talk.”
Zvi’s “asynchronous and on demand” is also a hugely important consideration for me.
This seems orthogonal to the broadcast vs. 1-on-1 model. E.g., email threads are a thing, as are comment threads that whose primary value is a dialog between two people.
I think one of the most useful things a public discussion can do is create more shared terminology, in order to facilitate future discussions, both private and public. Examples that I think have been clearly good (aside from, like, the whole of the Sequences): Moloch, Slack.
Broadcasting is good for transferring information from people with more valuable time to people with less valuable time—in those cases “using up the explainer’s time” much more than doubles the cost. I expect it would be better to do transmitting information 1-on-1 transmission to specialized explainers who broadcast it, but we aren’t good at that.
This suggests some models:
(thinker → conversation partner) repeated for everyone of the community
thinker ⇒ everyone in the community
thinker → explainer ⇒ everyone in the community
(where ‘=>’ implies telling a lot of people via an essay)
These models are in order of decreasing net cost to the system, if we assume the thinker’s time is the most expensive.
Let me propose another model
thinker ⇒ group of readers in a nearby inferential space ⇒ everyone in the community
I think there’s room for LW to intervene on this part of the model. Here are some ways to do that:
Create a new type of post which is an ‘explainer’ - something that does not claim original content, but does purport to re-explain another’s content in a more broadly understandable way. Give the best of these a permanent/prime position on the site (i.e. the way that R:AZ/Codex/HPMOR currently are) to incentivise their creation.
Let an author whitelist a set of other users, who may edit/re-write their posts, and who are rewarded for doing so in a way that successfully communicates the post to a wider range of people.
A small scale example of this at work would be how the commenters on Goodhart Taxonomy came up with multiple concrete examples using basketball, and these were incorporated into the post (which vastly improved my understanding of it, and I think significantly improved others’ ability to quickly grok the post).
Create a new type of sequence which is a ‘first-draft’, and create a system whereby an author can write a sequence of posts, get feedback from the comments, before re-writing / structuring the sequence, and potentially handing over the ‘next-draft’ to a different author.
This is not specifically for the original thinker, but can be thought of as analogy to how lecture series at universities work; the first lecturer of an advanced course may be the person who invented the ideas, but subsequent lecturers can be different individuals who nonetheless iterate on the explanations used.
This would go hand-in-hand with a feature whereby authors and trusted users can create exercises for posts, and those who successfully answer them can create new exercises themselves + mark others’ answers. Incentivise taking and creating such exercises. Use these as a test of the quality of an explanation.
By investing in really great explanations you can save total time compared to one-on-one explanations. I think a very small fraction of public discussion falls into this category / meets this bar.
Such discussion however, can be selected to be brought significantly into prominence; a site restructure towards focusing attention on such writing may be highly valuable.
Public discussion serves important non-information-transmission functions, e.g. building consensus, political maneuvering, norm enforcement, etc. In practice I think these are more important than information transmission.
While I broadly agree with the statement, the current top contender in my mind for why this is wrong, is a model whereby disagreement between key researchers in a field can be aided significantly by the other researchers, who hash out the lower-level details of the key researchers’ abstract intuitions.
An example of someone taking another’s ideas and implementing them concretely, is Eliezer’s A Human’s Guide to Words, being followed up by the application by Scott Alexander in the sequence Categorisation and Concepts. This helps to provide both a better understanding and a test of the abstract model. If two researchers disagree, others in their respective nearby inferential spaces can hash out many of the details. Naturally, this conversation can be aided by public discussion.
This is a model I’ve thought about but don’t yet feel I have strong evidence for.
The factor of 2 is a multiplier on the number of readers. So the number of readers matters. I think there are more lurkers on LW than people realize. See e.g. this obscure poll of mine that got over 100 votes from logged in users. It wouldn’t surprise me if I am speaking to an audience of over 50 people while typing in to this comment box.
An advantage of public discourse is that we have good methods for connecting readers with the posts that are likely to be most impactful for them. They can look at post titles, have posts suggested by friends, etc. Finding high value conversations in real life is more haphazard: networking, parties, small talk, etc.
The best approach might be a hybrid, where you test your ideas on a conversational audience and publish them when you’ve refined them enough that they’re digestible for a mass audience.
The factor of 2 is in the limit of infinitely many readers.
I agree that content discovery is tough. Though note that the question is largely about finding topics, not necessarily conversation partners, and here we have better mechanisms (if I’m talking with someone I know, I can bring up the topics most likely to interest them). I think that connecting people is a hard problem we’re not great at, though the intuition that it looks hard is also tied up with a few of the other problems (talking from higher to lower values of time, compensating talkers).
When something is communicated one-on-one, the cost is effectively doubled because it consumes the sender’s time. I think is roughly offset by a factor of 2 efficiency bonus for tailored/interactive talk relative to static public talk. Even aside from the efficiency of tailored talk, a factor of 2 isn’t very big compared to the other costs and benefits.
This feels to me like a strange model to apply here. I can see the logic of what you’re saying—if every “explanation event” requires time from an explainer and a receiver, then centralizing the explanations cuts the cost by only a factor of 2.
But it feels a bit to me like saying, “Everyone should write their own software. Each software use event requires time from a programmer and a user, so writing the programs once only cuts the cost by a factor of 2.”
Maybe the software analogy is unfair because any given user will use the same piece of software many times. But you could replace software with books, or movies or any other information product.
Broadcasting is good for transferring information from people with more valuable time to people with less valuable time—in those cases “using up the explainer’s time” much more than doubles the cost.
Maybe the way I would put it is: “Some people have much better things to say than others.”
If I have to get everything from one-on-one conversation, then I only have access to the thoughts of the people I can get access to in person.
So then, whether it’s worth it to do broadcasting would depend on:
1) How much does information content degrade if it’s passed around by word of mouth?
2) How much better are the best ideas in our community than median ideas?
3) How valuable is it for the median person to hear the best ideas?
In my envisioned world, the listener compensates the talker in some other way (if the talker is not sufficiently motivated by helping/influencing the listener). It won’t usually be the case that the talker can be compensated by taking a turn as the listener, unless the desired exchange of information happens to line up perfectly that day.
“How much does information content degrade if it’s passed around by word of mouth?” matters for the particular alternative strategy where you try to do a bucket brigade from people with more valuable time to people with less valuable time.
In custom-made-for-you software case, the programmer spends radically more time than the user, so rewriting the software for each user much more than doubles the cost.
In the custom-made-for-you talk case, the talker spends the same amount of time as the listener and so the effect is doubling.
Sometimes a static explanation is radically better than a tailored one-on-one interaction, instead of slightly worse. In that case I think broadcasting is very useful, just like for software. I think that’s unusual (for example, I expect that reading this comment is worse for you than covering this topic in a one-on-one discussion). Even in cases where a large amount of up-front investment can improve communication radically, I think the more likely optimal strategy is to spend a bunch of time creating an artifact that explains something well, and then to additionally discuss the issue one-on-one as a complement.
(This might look like an explanation of why there aren’t good public explanations of much of my research. That’s actually an independent issue though, I don’t endorse it based on these arguments.)
It’s easy to overestimate the intrinsic benefits of public discussion
Do you have an explanation for why people would systematically overestimate? (Or is this just an observation that it appears to you that many people do overestimate, so it must be easy to do :P)
I expect that much of the arguing in favor of maximum freedom for the commenter is motivated by memory of the Wild West-type areas of the internet, which some of its users have gotten used to as their safe space.
I think rationalists systematically overestimate the benefits of public discussion (and underestimate the costs).
When something is communicated one-on-one, the cost is effectively doubled because it consumes the sender’s time. I think is roughly offset by a factor of 2 efficiency bonus for tailored/interactive talk relative to static public talk. Even aside from the efficiency of tailored talk, a factor of 2 isn’t very big compared to the other costs and benefits.
So I think the advantages of public discussion are not mostly the efficiency from many people being able to read something that got written once. Instead they are things like:
One-on-one explanations often require the listener to compensate the explainer for their time. Right now we have no good institutions for this problem, so it relies on the explainer caring about what the listener believes or expecting to receive something in return.
By investing in really great explanations you can save total time compared to one-on-one explanations. I think a very small fraction of public discussion falls into this category / meets this bar.
Relative to the potential of both modes, we are currently worse (!) at one-on-one transmission than written broadcasting. For example, we have larger unnecessary overhead and political diversion from one-on-one transmission.
Broadcasting is good for transferring information from people with more valuable time to people with less valuable time—in those cases “using up the explainer’s time” much more than doubles the cost. I expect it would be better to do transmitting information 1-on-1 transmission to specialized explainers who broadcast it, but we aren’t good at that.
Often one-on-one discussion creates artifacts that are cheap to share, and at that point you might as well share (even though it’s a minority of the value and so is only worth doing if it doesn’t increase the cost much). For example, it’s extremely valuable to sort through the evidence and clarify our thinking, and once we do that we its relatively cheap to share. And discussions often produce transcripts that are worth sharing (though there is often political overhead from publishing a discussion, which often outweighs the benefits).
Public discussion serves important non-information-transmission functions, e.g. building consensus, political maneuvering, norm enforcement, etc. In practice I think these are more important than information transmission.
I do think that it’s great to iterate on and improve the rationalist crowd’s fora for public discussion (and I think use of FB is evidence for an unmet need). But:
It’s easy to overestimate the intrinsic benefits of public discussion and so be overly surprised when it doesn’t catch on. (Or to make overly strong inference about groups given the lack of useful public discussion.)
I expect there would be larger returns to improving our local institutions for one-on-one information transmission.
Given that rationalists estimate the value of public discourse (and also private discourse) higher than almost anyone else, it is certainly plausible that we overestimate it. Given I estimate it higher than most rationalists, it’s even more plausible that I overestimate it. But if anything, I think we underestimate it, and Paul points to some of the reasons why.
Paul is implicitly, I believe, employing a sort of ‘student-teacher’ model of communication; the teacher transmits communication, the student learns it. The customization and lack of politicization in person is about a factor of two better, which then roughly ‘cancels out’ the increased cost. But if that’s true, it’s then pointed out that there are a number of other key advantages to public discourse. Paul’s extra reasons seem super important, especially if we add creation of common knowledge and vocabulary to the last one. He also points out that teacher time is typically more valuable than student time, meaning the cost ratio is more than two, perhaps much more. And time spent writing dwarfs the time each person spends reading, so that further increases the ratio.
So it seems like even if Paul’s estimates are complete and right, we should still default to lots of public discourse?
To that I’d add at least:
Reference (we can cite the past as evidence slash explanation that allows us to build upon it, which is super valuable even when almost no one clicks on the link)
Error correction and feedback (if we do 1-on-1, and we get it wrong, we don’t get to fix mistakes or get help from others), review (we can go back and remind ourselves)
Selection (the best explanations and teachers get used more)
Opportunity (if something is being taught in private, I often won’t even know it’s a thing one could learn, at all)
Learning through writing (I learn a ton from having to write down my ideas robustly)
Better crafted explanations by taking the time to get them right and general
Asynchronous and on demand (I can teach today when I have time, or over two weeks as thoughts come to me, then people can learn when they need to learn and have time, which radically decreases effective cost) which only gets more important as people comment and converse slowly over time.
+1 on “learning through writing,” this factor alone is my main disagreement with Paul’s comment. For most people the biggest low-hanging improvement to their thinking might just be to write it down. Once it’s written down it’s not nearly that much effort to edit enough to post in a public discussion.
Zvi already mentioned this, but I just want to emphasize that I think error finding and correction (including noticing possible ambiguities and misunderstandings) is one of the most important functions of discussion, and it works much better in a group discussion (especially for hard to detect errors) because everyone interested in the discussion can see the error as soon as one person notices it and points it out. If you do a series of one-on-one discussions, to achieve the same effect you’d have to track down all the previous people you talked to and inform them. I think in practice very few people would actually bother to do that, due to subconscious status concerns among other reasons, which would lead to diverging epistemic states among people thinking about the topic and ensuing general confusion.(Sometimes an error is found that the “teacher” doesn’t recognize as an error due to various cognitive biases, which would make it impossible to track down all the previous discussants and inform them of the error.)
Zvi’s “asynchronous and on demand” is also a hugely important consideration for me. In an one-on-one discussion, the fact that I can ask questions and get immediate responses is often counterbalanced by the feeling that I don’t have time to think, digest an idea, and decide whether it makes sense or not, because the other person is waiting for me to nod so they can move on (and that person’s time is usually pretty valuable). So frequently I just nod and think that I’ll figure it out later on my own time, but that means I’m not actually taking advantage of the main benefit of having a one-on-one discussion. I think Paul might not be taking this into account when he says “a factor of 2 efficiency bonus for tailored/interactive talk relative to static public talk.”
This seems orthogonal to the broadcast vs. 1-on-1 model. E.g., email threads are a thing, as are comment threads that whose primary value is a dialog between two people.
I think one of the most useful things a public discussion can do is create more shared terminology, in order to facilitate future discussions, both private and public. Examples that I think have been clearly good (aside from, like, the whole of the Sequences): Moloch, Slack.
This suggests some models:
(thinker → conversation partner) repeated for everyone of the community
thinker ⇒ everyone in the community
thinker → explainer ⇒ everyone in the community
(where ‘=>’ implies telling a lot of people via an essay)
These models are in order of decreasing net cost to the system, if we assume the thinker’s time is the most expensive.
Let me propose another model
thinker ⇒ group of readers in a nearby inferential space ⇒ everyone in the community
I think there’s room for LW to intervene on this part of the model. Here are some ways to do that:
Create a new type of post which is an ‘explainer’ - something that does not claim original content, but does purport to re-explain another’s content in a more broadly understandable way. Give the best of these a permanent/prime position on the site (i.e. the way that R:AZ/Codex/HPMOR currently are) to incentivise their creation.
Let an author whitelist a set of other users, who may edit/re-write their posts, and who are rewarded for doing so in a way that successfully communicates the post to a wider range of people.
A small scale example of this at work would be how the commenters on Goodhart Taxonomy came up with multiple concrete examples using basketball, and these were incorporated into the post (which vastly improved my understanding of it, and I think significantly improved others’ ability to quickly grok the post).
Create a new type of sequence which is a ‘first-draft’, and create a system whereby an author can write a sequence of posts, get feedback from the comments, before re-writing / structuring the sequence, and potentially handing over the ‘next-draft’ to a different author.
This is not specifically for the original thinker, but can be thought of as analogy to how lecture series at universities work; the first lecturer of an advanced course may be the person who invented the ideas, but subsequent lecturers can be different individuals who nonetheless iterate on the explanations used.
This would go hand-in-hand with a feature whereby authors and trusted users can create exercises for posts, and those who successfully answer them can create new exercises themselves + mark others’ answers. Incentivise taking and creating such exercises. Use these as a test of the quality of an explanation.
Such discussion however, can be selected to be brought significantly into prominence; a site restructure towards focusing attention on such writing may be highly valuable.
While I broadly agree with the statement, the current top contender in my mind for why this is wrong, is a model whereby disagreement between key researchers in a field can be aided significantly by the other researchers, who hash out the lower-level details of the key researchers’ abstract intuitions.
An example of someone taking another’s ideas and implementing them concretely, is Eliezer’s A Human’s Guide to Words, being followed up by the application by Scott Alexander in the sequence Categorisation and Concepts. This helps to provide both a better understanding and a test of the abstract model. If two researchers disagree, others in their respective nearby inferential spaces can hash out many of the details. Naturally, this conversation can be aided by public discussion.
This is a model I’ve thought about but don’t yet feel I have strong evidence for.
The factor of 2 is a multiplier on the number of readers. So the number of readers matters. I think there are more lurkers on LW than people realize. See e.g. this obscure poll of mine that got over 100 votes from logged in users. It wouldn’t surprise me if I am speaking to an audience of over 50 people while typing in to this comment box.
An advantage of public discourse is that we have good methods for connecting readers with the posts that are likely to be most impactful for them. They can look at post titles, have posts suggested by friends, etc. Finding high value conversations in real life is more haphazard: networking, parties, small talk, etc.
The best approach might be a hybrid, where you test your ideas on a conversational audience and publish them when you’ve refined them enough that they’re digestible for a mass audience.
The factor of 2 is in the limit of infinitely many readers.
I agree that content discovery is tough. Though note that the question is largely about finding topics, not necessarily conversation partners, and here we have better mechanisms (if I’m talking with someone I know, I can bring up the topics most likely to interest them). I think that connecting people is a hard problem we’re not great at, though the intuition that it looks hard is also tied up with a few of the other problems (talking from higher to lower values of time, compensating talkers).
This feels to me like a strange model to apply here. I can see the logic of what you’re saying—if every “explanation event” requires time from an explainer and a receiver, then centralizing the explanations cuts the cost by only a factor of 2.
But it feels a bit to me like saying, “Everyone should write their own software. Each software use event requires time from a programmer and a user, so writing the programs once only cuts the cost by a factor of 2.”
Maybe the software analogy is unfair because any given user will use the same piece of software many times. But you could replace software with books, or movies or any other information product.
I think something like this is a crux for me:
Maybe the way I would put it is: “Some people have much better things to say than others.”
If I have to get everything from one-on-one conversation, then I only have access to the thoughts of the people I can get access to in person.
So then, whether it’s worth it to do broadcasting would depend on:
1) How much does information content degrade if it’s passed around by word of mouth?
2) How much better are the best ideas in our community than median ideas?
3) How valuable is it for the median person to hear the best ideas?
In my envisioned world, the listener compensates the talker in some other way (if the talker is not sufficiently motivated by helping/influencing the listener). It won’t usually be the case that the talker can be compensated by taking a turn as the listener, unless the desired exchange of information happens to line up perfectly that day.
“How much does information content degrade if it’s passed around by word of mouth?” matters for the particular alternative strategy where you try to do a bucket brigade from people with more valuable time to people with less valuable time.
In custom-made-for-you software case, the programmer spends radically more time than the user, so rewriting the software for each user much more than doubles the cost.
In the custom-made-for-you talk case, the talker spends the same amount of time as the listener and so the effect is doubling.
Sometimes a static explanation is radically better than a tailored one-on-one interaction, instead of slightly worse. In that case I think broadcasting is very useful, just like for software. I think that’s unusual (for example, I expect that reading this comment is worse for you than covering this topic in a one-on-one discussion). Even in cases where a large amount of up-front investment can improve communication radically, I think the more likely optimal strategy is to spend a bunch of time creating an artifact that explains something well, and then to additionally discuss the issue one-on-one as a complement.
Yes, but in part that’s because being in your company is such a joy. ;-)
(This might look like an explanation of why there aren’t good public explanations of much of my research. That’s actually an independent issue though, I don’t endorse it based on these arguments.)
Do you have an explanation for why people would systematically overestimate? (Or is this just an observation that it appears to you that many people do overestimate, so it must be easy to do :P)
Mostly introspective.
I expect that much of the arguing in favor of maximum freedom for the commenter is motivated by memory of the Wild West-type areas of the internet, which some of its users have gotten used to as their safe space.