Draft/WIP: The Working Memory Hypothesis re: Intellectual Progress
Strong claim, medium felt
So I’m working with the hypothesis that working memory (or something related) is a major bottleneck on progress within a given field. This has implications on what sort of things fields need.
Basic idea is that you generally need to create new concepts out of existing sub-concepts. You can only create a concept if you can hold the requisite sub-concepts in your head at once. Default working memory limits is 4-7 chunks. You can expand that somewhat by writing things down on whiteboards. I’m not sure exactly what the practical limits are.
But there is some upper limit on how complex a concept people can work with, even the smartest people in the field. So there are some ideas you just can’t invent, if they require 30 moving parts to consider at once. If an idea has 30 moving parts, it’s necessary to find some way to conceptualize it as having fewer parts.
Fortunately, this is possible. When I first learn a new concept, it’s unfamiliar, and I have to track all of it’s individual pieces separately – if it has 5 sub-concepts it takes 5 chunks of working memory.
But eventually I learn to think of it as a single concept, and then it only takes 1.
So, an important aspect of a field is distillation – improving explanations to help people think of clusters-of-ideas as a single concept they can reason about as if it took up only one chunk.
Distillation is a sort of tedious process though. So part of why Bell Labs is a great idea is that you can have all these specialists of different fields working in the same building, and whenever one of them wants to learn something they can just walk down the hall and ask the guy who invented it “hey, how does this work”, and they get an explanation. And if they don’t understand it fully, they can immediately, frictionlessly, ask followup questions.
This doesn’t just help them understand new ideas – it reduces those ideas in conceptual complexity until they only take up a single chunk, enabling them to be combined with other ideas.
This has implications for the AI Alignment field – investing in distillation and co-location are both potentially quite valuable. (But, correspondingly – watch out for things that improve the distillation of ideas that feed into AI capabilities)
Yuppers. Yeah, the idea I’m trying to get at here could be conceptualized as “take the underlyinggenerator that outputs Research Debt, and then lean hard into using it as an explanatory theory, and see that other hypotheses turn up when you take that seriously.”
(I’d already read research debt too at the time Oli first explained this concept to me. I think Oli’s additional contribution was thinking in terms of chunks being a limiting factor. He didn’t specific working memory precisely as the constraint. I later thought about the intersection of working-memory-in-particular after writing You Have About Five Words and later thinking about some implications on this comment here)
Oli had left the number of chunks available deliberately vague, and I’m now concretely predicting that people can only build theories systems that don’t require them to hold more than 4-10* chunks at once.
*where “10” is an ass-pulled number for “how much your working memory can really be improved via writing things done.”
[I don’t know if Oli thinks working-memory-in-particular makes sense to think of as the bottleneck]
After learning a new concept, it is important to “play with it” for a while. Because the new concept is initially not associated with anything, so you probably will not see what it is good for.
For example, if someone tells you “a prime number is an integer number greater than one that can only be divided by itself and by one”, that is easy to understand (even easier if they also give you a few examples of primes and non-primes), but it is not obvious why is this concept important and how could it be used.
But when the person also tells you “the number of primes is infinite… each integer can be uniquely factored into primes… some numbers are obviously not primes, but we don’t know a simple method to find out whether a large number is a prime… in arithmetic modulo n you can define addition, subtraction, and multiplication for any n, but you can unambiguously define division only when n is prime...” and perhaps introduces a concept of “relative primes” and the Chinese remainder theorem… then you may start getting ideas of how it could be useful, such as “so, if we take two primes so big that we can barely verify their primeness, and multiply them, it will be almost impossible to factor the result, but it would be trivial to verify when the original two numbers are provided—I wonder whether we could use this as a form of signature.”
Draft/WIP: The Working Memory Hypothesis re: Intellectual Progress
Strong claim, medium felt
So I’m working with the hypothesis that working memory (or something related) is a major bottleneck on progress within a given field. This has implications on what sort of things fields need.
Basic idea is that you generally need to create new concepts out of existing sub-concepts. You can only create a concept if you can hold the requisite sub-concepts in your head at once. Default working memory limits is 4-7 chunks. You can expand that somewhat by writing things down on whiteboards. I’m not sure exactly what the practical limits are.
But there is some upper limit on how complex a concept people can work with, even the smartest people in the field. So there are some ideas you just can’t invent, if they require 30 moving parts to consider at once. If an idea has 30 moving parts, it’s necessary to find some way to conceptualize it as having fewer parts.
Fortunately, this is possible. When I first learn a new concept, it’s unfamiliar, and I have to track all of it’s individual pieces separately – if it has 5 sub-concepts it takes 5 chunks of working memory.
But eventually I learn to think of it as a single concept, and then it only takes 1.
So, an important aspect of a field is distillation – improving explanations to help people think of clusters-of-ideas as a single concept they can reason about as if it took up only one chunk.
Distillation is a sort of tedious process though. So part of why Bell Labs is a great idea is that you can have all these specialists of different fields working in the same building, and whenever one of them wants to learn something they can just walk down the hall and ask the guy who invented it “hey, how does this work”, and they get an explanation. And if they don’t understand it fully, they can immediately, frictionlessly, ask followup questions.
This doesn’t just help them understand new ideas – it reduces those ideas in conceptual complexity until they only take up a single chunk, enabling them to be combined with other ideas.
This has implications for the AI Alignment field – investing in distillation and co-location are both potentially quite valuable. (But, correspondingly – watch out for things that improve the distillation of ideas that feed into AI capabilities)
This seems highly related to Chris Olah’s Research Debt.
(That was indeed the piece that crystallized this intuition for me, and I think Ray got this broader concept from me)
Yuppers. Yeah, the idea I’m trying to get at here could be conceptualized as “take the underlying generator that outputs Research Debt, and then lean hard into using it as an explanatory theory, and see that other hypotheses turn up when you take that seriously.”
(I’d already read research debt too at the time Oli first explained this concept to me. I think Oli’s additional contribution was thinking in terms of chunks being a limiting factor. He didn’t specific working memory precisely as the constraint. I later thought about the intersection of working-memory-in-particular after writing You Have About Five Words and later thinking about some implications on this comment here)
Oli had left the number of chunks available deliberately vague, and I’m now concretely predicting that people can only build theories systems that don’t require them to hold more than 4-10* chunks at once.
*where “10” is an ass-pulled number for “how much your working memory can really be improved via writing things done.”
[I don’t know if Oli thinks working-memory-in-particular makes sense to think of as the bottleneck]
After learning a new concept, it is important to “play with it” for a while. Because the new concept is initially not associated with anything, so you probably will not see what it is good for.
For example, if someone tells you “a prime number is an integer number greater than one that can only be divided by itself and by one”, that is easy to understand (even easier if they also give you a few examples of primes and non-primes), but it is not obvious why is this concept important and how could it be used.
But when the person also tells you “the number of primes is infinite… each integer can be uniquely factored into primes… some numbers are obviously not primes, but we don’t know a simple method to find out whether a large number is a prime… in arithmetic modulo n you can define addition, subtraction, and multiplication for any n, but you can unambiguously define division only when n is prime...” and perhaps introduces a concept of “relative primes” and the Chinese remainder theorem… then you may start getting ideas of how it could be useful, such as “so, if we take two primes so big that we can barely verify their primeness, and multiply them, it will be almost impossible to factor the result, but it would be trivial to verify when the original two numbers are provided—I wonder whether we could use this as a form of signature.”