If the note is a burden, I’d say it is a problem of the note-taking system rather than of the note itself.
In my case it’s not the individual notes that are burdensome, it’s the sheer volume of notes that is the burden (“the dose makes the poison”). As the notes I have to recall, digest, and pass judgement on increases, so too does the cognitive load. The kind of notes I’m talking about should aid in making decisions or improving processes: more notes and dead-end paths slows down the decision making process. That’s robbing my future-self.
What’s interesting is preemptive “sorting” goes against one of the core tenants of Zettelkasten which promotes this idea of connections between atomic notes arising organically.
I am curious about Gwern’s Suggestion of what to do with lists one is not sure how to sort, but I cannot understand the suggestion.
I am curious about Gwern’s Suggestion of what to do with lists one is not sure how to sort, but I cannot understand the suggestion.
I am just guessing here, but it’s probably something like this: Find a way to calculate the “distance” between the items (in case of notes, perhaps if they link to each other, they are “closer”; or maybe also if they have the same keywords, dunno), then create an N-dimensional structure out of that, then find the longest dimension in that structure and project all nodes to that axis, and now you have a linear ordering that feels somewhat meaningful. (Probably not exactly this.)
As the notes I have to recall, digest, and pass judgement on increases, so too does the cognitive load.
Yeah. In my opinion, an ideal system should somehow take into account that some notes are old, unreviewed, etc., and give those notes lower priority, e.g. they would be displayed in a gray color and come later in search results. So they mostly wouldn’t bother you, unless you find them by a hyperlink, or you use the right keywords in search.
I haven’t tested this idea in practice. It’s just based on my intuition of how human memory works. I mean, with humans we don’t have this feeling that learning about new things is mostly a cognitive debt. (Or do we? I think a similar sentiment was expressed in a Sherlock Holmes book, where the genius detective said something like “I don’t want to know that the Earth is round, because that is irrelevant to my work”.) That’s because our memories are connected by associations, and the frequently recalled ones are more likely to come to mind. Maybe we could capture this intuition in an algorithm.
Yes, although more concretely it would be to use an off-the-shelf neural embedding, then cosine distance, and then TSP (or one of the many other algorithms) to find a linear ordering which minimizes the total distance. I never got around to TSP or anything fancy because I discovered for my use case, the dumbest possible greedy algorithm (define the starting point as the newest entry; take the nearest entry as the second entry; repeat until you run out) worked well, so I never got around to seeing if the fancier ones delivered a visible improvement.
I don’t know how well exactly that would work in a Zettelkasten context with a lot of notes, rather than a relatively limited set of curated notes. A seriation approach which works nicely at n = 20 might perform very badly at n = 20,000, if only because a list of 20,000 items is hard to make useful.
My off the cuff suggestion would be that you would want to instead initialize a tag-category hierarchy and work within it from then on: create a tag-category hierarchy automatically, and seriate within that. Something like, do k-means clustering of the 20,000 and break them up into 10 clusters; label with a LLM; break each cluster up into 10 sub-clusters; label; and so on until you have 10-100 notes in each cluster (so maybe 4 levels), and then seriate within the lowest level clusters. The expert human then can superimpose some additional metadata and rewrite names etc to improve the inferred taxonomy. Only the most pathologically obsessed Zettelkasten user would ever do it by hand, but even a noisy and inapt breakdown is a lot better than a flat list of 20,000 (no mater how well seriated that is)! See https://gwern.net/design#sort-by-magic and https://gwern.net/design#future-tag-features
Yeah. In my opinion, an ideal system should somehow take into account that some notes are old, unreviewed, etc., and give those notes lower priority, e.g. they would be displayed in a gray color and come later in search results. So they mostly wouldn’t bother you, unless you find them by a hyperlink, or you use the right keywords in search.
Oh cool. You got me thinking and I might walk-back from what i said earlier—maybe individual notes are a burden? For example, let’s say I’m working on a title sequence for a sitcom, specifically it’s a Dark Comedy about an Oncologist. And I need to make some decisions about what style, what content etc. etc.. So I type “title sequence” or “opening credits” into my note taking system and it comes up with the example from my OP:
“Castle Bryant Johnston are the firm that did the opening titles of Cheers”
Imagine that’s the note in full. And it’s top of the heap. So I go to youtube and watch the opening titles of Cheers, and other television shows the firm worked on. (Adding further to the burden: maybe I make some more notes, or observations. Notes create notes—when does it end?)
But it still doesn’t actually resolve the decision: what this Dark Comedy about an Oncologist title sequence should be?
See, all the unreviewed old, low priority notes could be greyed out. But does it solve the problem, does make the decision any faster? No. that defeats the entire purpose of writing the note, because a note is work that is speculating on a payoff. Too many of these (as is the case with me) and the whole pattern is useless.
Through having written all this out in a reply, it seems to me the solution is simply: I should instead speculate on why they opted for historical photographs rather than write the factoid- i.e. the note should speculate on what was the reasoning behind that choice. Producing a note more like:
Castle Bryant and Johnston probably opted to make the opening titles of Cheers with static, historical photographs, which serves as a macrocosm for the lives of the main cast, who are stuck in a routine of coming to the bar—Sam in particularly clinging to former glories
And even better note, I would make observations from other Castle Bryant and Johnston title sequences and speculate on why they opted for the creative decisions they did—before I commit the note to my file!
The action and location, around the table of Roseanne shows the family dynamics literally around the Matriarch like a working class Queen’s court...blah blah blah while the single unbroken camera shot produces a feeling of… rah rah rah
I think this is better because if I make these observations at the time I write the note, rather than when I recall the note, I’ve shifted the work from my future-self to the past, and it provide heuristics or ways to think about the Oncologist Sitcom decision (or other, similar, decisions!).
So individual notes can be a burden if they don’t provide readymade answers. The bottleneck isn’t searching or recall for relevance—if the notes are too spartan they aren’t useful even if they are relevant. The issue is providing ways of making decisions and doing things well: and one way to speed up the process is by front-loading speculating on why a decision or action was taken. Rather than taking a spartan note.
Associations are fine—but what’s more important (in my view) is heuristics, rules, and techniques that are immediately applicable.
If the note is a burden, I’d say it is a problem of the note-taking system rather than of the note itself.
(That said, I think it is possible that all existing systems suck, and we need to invent something much better.)
In my case it’s not the individual notes that are burdensome, it’s the sheer volume of notes that is the burden (“the dose makes the poison”). As the notes I have to recall, digest, and pass judgement on increases, so too does the cognitive load. The kind of notes I’m talking about should aid in making decisions or improving processes: more notes and dead-end paths slows down the decision making process. That’s robbing my future-self.
What’s interesting is preemptive “sorting” goes against one of the core tenants of Zettelkasten which promotes this idea of connections between atomic notes arising organically.
I am curious about Gwern’s Suggestion of what to do with lists one is not sure how to sort, but I cannot understand the suggestion.
I am just guessing here, but it’s probably something like this: Find a way to calculate the “distance” between the items (in case of notes, perhaps if they link to each other, they are “closer”; or maybe also if they have the same keywords, dunno), then create an N-dimensional structure out of that, then find the longest dimension in that structure and project all nodes to that axis, and now you have a linear ordering that feels somewhat meaningful. (Probably not exactly this.)
Yeah. In my opinion, an ideal system should somehow take into account that some notes are old, unreviewed, etc., and give those notes lower priority, e.g. they would be displayed in a gray color and come later in search results. So they mostly wouldn’t bother you, unless you find them by a hyperlink, or you use the right keywords in search.
I haven’t tested this idea in practice. It’s just based on my intuition of how human memory works. I mean, with humans we don’t have this feeling that learning about new things is mostly a cognitive debt. (Or do we? I think a similar sentiment was expressed in a Sherlock Holmes book, where the genius detective said something like “I don’t want to know that the Earth is round, because that is irrelevant to my work”.) That’s because our memories are connected by associations, and the frequently recalled ones are more likely to come to mind. Maybe we could capture this intuition in an algorithm.
Yes, although more concretely it would be to use an off-the-shelf neural embedding, then cosine distance, and then TSP (or one of the many other algorithms) to find a linear ordering which minimizes the total distance. I never got around to TSP or anything fancy because I discovered for my use case, the dumbest possible greedy algorithm (define the starting point as the newest entry; take the nearest entry as the second entry; repeat until you run out) worked well, so I never got around to seeing if the fancier ones delivered a visible improvement.
I don’t know how well exactly that would work in a Zettelkasten context with a lot of notes, rather than a relatively limited set of curated notes. A seriation approach which works nicely at n = 20 might perform very badly at n = 20,000, if only because a list of 20,000 items is hard to make useful.
My off the cuff suggestion would be that you would want to instead initialize a tag-category hierarchy and work within it from then on: create a tag-category hierarchy automatically, and seriate within that. Something like, do k-means clustering of the 20,000 and break them up into 10 clusters; label with a LLM; break each cluster up into 10 sub-clusters; label; and so on until you have 10-100 notes in each cluster (so maybe 4 levels), and then seriate within the lowest level clusters. The expert human then can superimpose some additional metadata and rewrite names etc to improve the inferred taxonomy. Only the most pathologically obsessed Zettelkasten user would ever do it by hand, but even a noisy and inapt breakdown is a lot better than a flat list of 20,000 (no mater how well seriated that is)! See https://gwern.net/design#sort-by-magic and https://gwern.net/design#future-tag-features
Oh cool. You got me thinking and I might walk-back from what i said earlier—maybe individual notes are a burden? For example, let’s say I’m working on a title sequence for a sitcom, specifically it’s a Dark Comedy about an Oncologist. And I need to make some decisions about what style, what content etc. etc.. So I type “title sequence” or “opening credits” into my note taking system and it comes up with the example from my OP:
Imagine that’s the note in full. And it’s top of the heap. So I go to youtube and watch the opening titles of Cheers, and other television shows the firm worked on. (Adding further to the burden: maybe I make some more notes, or observations. Notes create notes—when does it end?)
But it still doesn’t actually resolve the decision: what this Dark Comedy about an Oncologist title sequence should be?
See, all the unreviewed old, low priority notes could be greyed out. But does it solve the problem, does make the decision any faster? No. that defeats the entire purpose of writing the note, because a note is work that is speculating on a payoff. Too many of these (as is the case with me) and the whole pattern is useless.
Through having written all this out in a reply, it seems to me the solution is simply: I should instead speculate on why they opted for historical photographs rather than write the factoid- i.e. the note should speculate on what was the reasoning behind that choice. Producing a note more like:
And even better note, I would make observations from other Castle Bryant and Johnston title sequences and speculate on why they opted for the creative decisions they did—before I commit the note to my file!
I think this is better because if I make these observations at the time I write the note, rather than when I recall the note, I’ve shifted the work from my future-self to the past, and it provide heuristics or ways to think about the Oncologist Sitcom decision (or other, similar, decisions!).
So individual notes can be a burden if they don’t provide readymade answers. The bottleneck isn’t searching or recall for relevance—if the notes are too spartan they aren’t useful even if they are relevant. The issue is providing ways of making decisions and doing things well: and one way to speed up the process is by front-loading speculating on why a decision or action was taken. Rather than taking a spartan note.
Associations are fine—but what’s more important (in my view) is heuristics, rules, and techniques that are immediately applicable.