Also - the speed and quality improvements when having to redo homework lost to an undiscerning canine companion is also a corollary of this.
Perhaps the time it takes to 'redo' is a better measure than last mile - it's the entire effort, minus the initial solution-space exploration?
Just map quality q to e^q or something and it will be sublinear again.
Or more directly, if your argument for why effort scales linearly with perceived quality doesn't discuss how we perceive quality then something is wrong.
A more direct argument would be that it takes roughly an equal amount of effort to halve the distance from a rough work to its ideal. Going from 90% to 99% takes the same as going from 99% to 99.9% but the latter only covers a tenth of the distance. If our perception is more sensitive to the _absolute_ size of the error you get an exponential effort to improve something.
Your first line assumes that `q` fails to refer to an objective property. The `e^q` space isn't quality, as much as `e^t` isnt temperature (holding the property we are talking about fixed). Thus the comment ends up being circular.
The issue was with the word "it". In the sentence, that word is acting as an indirection to both q and e^q instead of referring to a unitary thing. So yes, "it" does become linear/sublinear, but "it" is no longer the original subject of discussion.
How often is a drawing really trashed and restarted?
There's the saying, "Plan to throw one away," but seems like it varies in practice (for software).
There are even books about patching paintings, like Master Disaster: Five Ways to Rescue Desparate Watercolors.
In architecture, it's understood the people, vehicles, and landscape are not as exact as the building or structure, and books encourage reusing magazine clippings, overhead projectors, and copy machines to generally "be quick" on execution.
Would like to see thoughts on comparing current process with the "Draw 50" series, where most of the skeleton is on paper by the first step, but the last is really the super-detailed, totally refined, owl.
From my very limited experience with art, it's more often the case that a work in progress creation is abandoned and then taken a stab at anew later than trashed and restarted. Or it is iterated on to a degree that it is not meaningfully different from a full restart.
I have a bit more experience with software and the only reason for why we don't plan throw one away is because it costs more money and the market pressure on software quality is too low to make stakeholders care. In my personal hobby coding, I often practice this (or do what I described above with art which is closer to abandoning until inspiration strikes again at which point a blank slate is more inviting). The closest thing professionally I get is a "spike" where I explore something via code with the output not being the code itself, but the knowledge attained which then becomes an input to new code writing.
So i can only speak from my own experience of the last 5 years of trying (and often failing!) to accurately copy or otherwise create various drawings.
Very rarely do I start completely from scratch, but usually adjust the drawing so much that maybe I should have. I wonder if I tracked the adjustments if I would find every line was redrawn in some cases.
Thing is, it is hard to see what part is 'off' until most of the other parts are right. Especially with highly symmetric drawings, where symmetries appear gradually as the whole thing comes together.
I believe that last-mile edits do not significantly improve the quality of (most) creative work. To produce high-quality work, one must have already "cached" their "motor heuristics," which, in simpler terms, means having dedicated thousands of hours to deep and deliberate practice in their field.
The definition of 'last-mile edits' is very subjective, though. If you're dealing with open systems, it's almost unthinkable to design something and not need to iterate on it until the desired outcome is achieved. In other domains, for example, playing an instrument, your skills need to have been honed previously: there's nothing that will make you sound better (without resorting to editing it electronically).
Judging by the comments here, I'm the only one, but I have no idea what he's talking about. Even the abstract:
> The act of creation is fractal exploration–exploitation under optimal feedback control. When resolution increases the portion of parameter space that doesn't make the artifact worse (acceptance volume) collapses. Verification latency and rate–distortion combine into a precision tax that scales superlinearly with perceived quality.
Is this just saying that it's ok if doodles aren't good, but the closer you get to the finished work, the better it has to be? If your audience can't understand what the hell you're talking about for simple ideas, you've gone too far.
The abstract is some of the worst writing I've read in a while. Trying to sound so very smart while being incapable of getting your point across. This whole article reeks of pretentiousness.
Yeah, it came off as complete nonsense. If someone were talking to me like this in person, I'd probably start suspecting they were doing it to distract me while their friend was outside stealing my hubcaps.
No you have an equal number of options (minor and major are effectively transpositions/rotations...e.g. the chord progressions are "m dim M m m M M" for minor (m-minor, M-major, dim-diminished) chord progression, vs "M m m M M m dim" for major).
The post is likely getting to the point that, for english-speaking/western audiences at least, you are more likely to find songs written in C major, and thus they are more familiar and 'safer'. You _can_ write great songs in Em, but it's just a little less common, so maybe requires more work to 'fit into tastes'.
Perceived quality is relative. It's roughly linearly related to rank position along some dimension, but moving up in rank requires exponential effort due to competition.
I would be surprised if anyone perceives quality like that. Like, are you saying that in a situation where there are only two examples of some type of work, it is impossible to judge whether one is much better than the other, it is only possible to say that it's better? What makes you think it works like this?
On a related note I wrote a few “poems” using anagrams. The principle is simple: take a short phrase and have each line in the poem be an anagram of it. You can’t do this with just any phrase; the letters need to be reasonably well balanced for the target language so you can still form pronouns, key grammatical verbs (to be, to have, etc.), and some basic structure.
It becomes interesting once sentences span multiple lines and you start using little tactical tricks to keep syntax, semantics, and the overall argument coherent while respecting the anagram constraint.
Using an anagram generator is of course a first step, but the landscapes it offers are mostly desert: the vast majority of candidates are nonsense, and those that are grammatical are usually thematically off relative to what you’ve already written. And yet, if the repeated anagram phrase is chosen well, it doesn’t feel that hard to build long, meaningful sentences. Subjectively, the difficulty seems to scale roughly proportionally with the length of the poem, rather than quadratically and beyond.
There’s a nice connection here to Sample Space Reducing (SSR) processes. The act of picking letters from a fixed multiset to form words, and removing them as you go, is a SSR. So is sentence formation itself: each committed word constrains the space of acceptable continuations (morphology, syntax, discourse, etc.).
> Many such stochastic processes, especially those that are associated with complex systems, become more constrained as they unfold, meaning that their sample-space, or their set of possible outcomes, reduces as they age. We demonstrate that these sample-space reducing (SSR) processes necessarily lead to Zipf’s law in the rank distributions of their outcomes.
> We note that SSR processes and nesting are deeply connected to phase-space collapse in statistical physics [21, 30–32], where the number of configurations does not grow exponentially with system size (as in Markovian and ergodic systems), but grows sub-exponentially. Sub-exponential growth can be shown to hold for the phase-space growth of the SSR sequences introduced here. In conclusion we believe that SSR processes provide a new alternative view on the emergence of scaling in many natural, social, and man-made systems.
In my case there are at least two coupled SSRs: (1) the anagrammatic constraint at the line level (letters being consumed), and (2) the layered SSRs of natural language that govern what counts as a well-formed and context-appropriate continuation (from morphology and syntax up through discourse and argumentation). In practice I ended up exploiting this coupling: by reserving or spending strategic words (pronouns, conjunctions, or semantically heavy terms established earlier), I could steer both the unfolding sentence and the remaining letter pool, and explore the anagram space far more effectively than a naive generator.
Very hand-wavy hypothesis: natural language is a complex, multi-layered SSR engine that happens to couple extremely well to other finite SSR constraints. That makes it unusually good at “solving” certain bounded combinatorial puzzles from the inside—up to and including, say, assembling IKEA furniture.
One extra nuance here: in the anagrammatic setting, the coupling between constraints is constitutive rather than merely referential. The same finite multiset of letters simultaneously supports the combinatorial constraint (what strings are formable) and the linguistic constraint (what counts as a syntactically and discursively acceptable move), so every choice is doubly binding. That’s different from cases like following IKEA instructions, where language operates as an external controller that refers to another state space (parts, tools, assembly steps) without sharing its “material” degrees of freedom. This makes the anagram case feel like a toy model where syntax and semantics are not two separate realms but two intertwined SSR layers over one shared substrate—suggesting that what we call “reference” might itself be an emergent pattern in how such nested SSR systems latch onto each other.
Perhaps a controversial view on this particular forum but I find the tendency of a certain type of person* to write about everything in this overly-technical way regardless of whether it is appropriate to the subject matter to be very tiresome ("executing cached heuristics", "constrained the search space").
*I associate it with the asinine contemporary "rationalist" movement (LessWrong et al.) but I'm not making any claims the author is associated with this.
I think it's a trick. It seems to be the article is just a series of ad-hoc assumptions and hypotheses without any support. The language aims to hide this, and makes you think about the language instead of its contents. Which is logically unsound: In a sharp peak, micro optimizations would give you a clearer signal where the optimum lies since the gradient is steeper.
> In a sharp peak, micro optimizations would give you a clearer signal where the optimum lies since the gradient is steeper.
I would refuse to even engage with the piece on this level, since it lends credibility to the idea that the creative process is even remotely related to or analogous to gradient descent.
I wouldn't jump to call it a trick, but I agree, the author sacrificed too much clarity in a try for efficiency.
The author set up an interesting analogy but failed to explore where it breaks down or how all the relationships work in the model.
My inference about the author's meaning was such: In a sharp peak, searching for useful moves is harder because you have fewer acceptable options as you approach the peak.
Fewer absolute or relative? If you scale down your search space... This only makes some kind of sense if your step size is fixed. While I agree with another poster that a reduction of a creative process to gradient descent is not wise, the article also misses the point what makes such a gradient descent hard -- it's not sharp peaks, it's the flat area around them -- and the presence of local minima.
It's a middle school essay that is trying to score points based on the number of metaphors used. Very unappealing and I wouldn't call it technical.
EDIT: For all the people saying the writing is inspired by math/cs, that's not at all true. That's not how technical writing is done. This guy is just a poser.
A bit harsh, but I see what you mean. It is tempting to try and fit every description of the world into a rigorous technical straightjacket, perhaps because it feels like you have understood it better?
Maybe it is similar to how scientist get flack for writing in technical jargon instead of 'plain language'. Partly it is a necessity - to be unambiguous - however it is also partly a choice, a way to signal that you are doing Science, not just describing messing about with chemicals or whatever.
I have observed it too, it is heavily inspired by economics and mathematics.
Saying "it's better to complete something imperfect than spend forever polishing" - dull, trite, anyone knows that. Saying "effort is a utility curve function that must be clamped to achieve meta-optimisation" - now that sounds clever
If I was going to be uncharitable, I think there is are corners of the internet where people write straightforward things dressed it up in technical language to launder it as somehow academic and data driven.
And you're right, it does show up in the worse parts of the EA / rationalist community.
(This writing style, at its worst, allows people to say things like "I don't want my tax spent on teaching poor kids to read" but without looking like complete psychopaths - "aggregate outcomes in standardised literacy programmes lag behind individualised tutorials")
That's not what the blog post here is doing, but it is definitely bad language use that is doing more work to obscure ideas than illuminate them
no, we need more of this, the opposite of this is Robin Williams destroying the poetry theory book in dead poeta society, the result was weak kids and one of them commited suicide. More technical stuff in relation to art is a good thing, but its expected that anglosaxon people have allergy to this, they think is somehow socialist or something and they need art to be unfefined etc
Respectfully, I have no idea what you're talking about. Dead Poets Society is a story and the message of the story isn't that Robin Williams' character is bad.
Are you saying my perspective is anti-socialist? What is "refined" art?
I appreciate this post as I think too many folks focus on the end before understanding what made it there. It's kind of asking what's the movie about before watching it or especially movie trailers that essentially shows way too much.
We should all take some time to better understand what brought us here to be better prepared for general creative work and uniqueness in the future...
lol I cited this exact scene as an example of typical anglosaxon conception of art, now you are crying that art has become shit but any attempt at scientific analysis is taken as a joke when actual poetry is even harder than Code, the amount of data you can compress on a single Word and rhyimes and stuff IS the hardest thing ever, but because you dont want to think someone can do an effort you want the Robin Williams and Dead Poets society to win and make art non scientifically understandable to anyone, if you cant do scientific or technical analysis of art thats your opinión but why the obsession on trashing anyone Who does It?
Also - the speed and quality improvements when having to redo homework lost to an undiscerning canine companion is also a corollary of this. Perhaps the time it takes to 'redo' is a better measure than last mile - it's the entire effort, minus the initial solution-space exploration?
Just map quality q to e^q or something and it will be sublinear again.
Or more directly, if your argument for why effort scales linearly with perceived quality doesn't discuss how we perceive quality then something is wrong.
A more direct argument would be that it takes roughly an equal amount of effort to halve the distance from a rough work to its ideal. Going from 90% to 99% takes the same as going from 99% to 99.9% but the latter only covers a tenth of the distance. If our perception is more sensitive to the _absolute_ size of the error you get an exponential effort to improve something.
Your first line assumes that `q` fails to refer to an objective property. The `e^q` space isn't quality, as much as `e^t` isnt temperature (holding the property we are talking about fixed). Thus the comment ends up being circular.
The issue was with the word "it". In the sentence, that word is acting as an indirection to both q and e^q instead of referring to a unitary thing. So yes, "it" does become linear/sublinear, but "it" is no longer the original subject of discussion.
How often is a drawing really trashed and restarted?
There's the saying, "Plan to throw one away," but seems like it varies in practice (for software).
There are even books about patching paintings, like Master Disaster: Five Ways to Rescue Desparate Watercolors.
In architecture, it's understood the people, vehicles, and landscape are not as exact as the building or structure, and books encourage reusing magazine clippings, overhead projectors, and copy machines to generally "be quick" on execution.
Would like to see thoughts on comparing current process with the "Draw 50" series, where most of the skeleton is on paper by the first step, but the last is really the super-detailed, totally refined, owl.
From my very limited experience with art, it's more often the case that a work in progress creation is abandoned and then taken a stab at anew later than trashed and restarted. Or it is iterated on to a degree that it is not meaningfully different from a full restart.
I have a bit more experience with software and the only reason for why we don't plan throw one away is because it costs more money and the market pressure on software quality is too low to make stakeholders care. In my personal hobby coding, I often practice this (or do what I described above with art which is closer to abandoning until inspiration strikes again at which point a blank slate is more inviting). The closest thing professionally I get is a "spike" where I explore something via code with the output not being the code itself, but the knowledge attained which then becomes an input to new code writing.
So i can only speak from my own experience of the last 5 years of trying (and often failing!) to accurately copy or otherwise create various drawings.
Very rarely do I start completely from scratch, but usually adjust the drawing so much that maybe I should have. I wonder if I tracked the adjustments if I would find every line was redrawn in some cases.
Thing is, it is hard to see what part is 'off' until most of the other parts are right. Especially with highly symmetric drawings, where symmetries appear gradually as the whole thing comes together.
I believe that last-mile edits do not significantly improve the quality of (most) creative work. To produce high-quality work, one must have already "cached" their "motor heuristics," which, in simpler terms, means having dedicated thousands of hours to deep and deliberate practice in their field.
The definition of 'last-mile edits' is very subjective, though. If you're dealing with open systems, it's almost unthinkable to design something and not need to iterate on it until the desired outcome is achieved. In other domains, for example, playing an instrument, your skills need to have been honed previously: there's nothing that will make you sound better (without resorting to editing it electronically).
A teacher told me once that editing poetry is like trying to open a glass jar. Eventually, you have to set it down or you’ll break the thing.
Judging by the comments here, I'm the only one, but I have no idea what he's talking about. Even the abstract:
> The act of creation is fractal exploration–exploitation under optimal feedback control. When resolution increases the portion of parameter space that doesn't make the artifact worse (acceptance volume) collapses. Verification latency and rate–distortion combine into a precision tax that scales superlinearly with perceived quality.
Is this just saying that it's ok if doodles aren't good, but the closer you get to the finished work, the better it has to be? If your audience can't understand what the hell you're talking about for simple ideas, you've gone too far.
The abstract is some of the worst writing I've read in a while. Trying to sound so very smart while being incapable of getting your point across. This whole article reeks of pretentiousness.
Yeah, it came off as complete nonsense. If someone were talking to me like this in person, I'd probably start suspecting they were doing it to distract me while their friend was outside stealing my hubcaps.
Very funny to put a bibtex citation under such a small piece of work
I liked the post but can someone explain how macro choices change the acceptance volume?
Is it their effect on the total number of available choices?
Does picking E minor somehow give you fewer options than C major (I'm not a musician)?
No you have an equal number of options (minor and major are effectively transpositions/rotations...e.g. the chord progressions are "m dim M m m M M" for minor (m-minor, M-major, dim-diminished) chord progression, vs "M m m M M m dim" for major).
The post is likely getting to the point that, for english-speaking/western audiences at least, you are more likely to find songs written in C major, and thus they are more familiar and 'safer'. You _can_ write great songs in Em, but it's just a little less common, so maybe requires more work to 'fit into tastes'.
edit: changed 'our' to english/western audiences
Perceived quality is relative. It's roughly linearly related to rank position along some dimension, but moving up in rank requires exponential effort due to competition.
I would be surprised if anyone perceives quality like that. Like, are you saying that in a situation where there are only two examples of some type of work, it is impossible to judge whether one is much better than the other, it is only possible to say that it's better? What makes you think it works like this?
On a related note I wrote a few “poems” using anagrams. The principle is simple: take a short phrase and have each line in the poem be an anagram of it. You can’t do this with just any phrase; the letters need to be reasonably well balanced for the target language so you can still form pronouns, key grammatical verbs (to be, to have, etc.), and some basic structure.
It becomes interesting once sentences span multiple lines and you start using little tactical tricks to keep syntax, semantics, and the overall argument coherent while respecting the anagram constraint.
Using an anagram generator is of course a first step, but the landscapes it offers are mostly desert: the vast majority of candidates are nonsense, and those that are grammatical are usually thematically off relative to what you’ve already written. And yet, if the repeated anagram phrase is chosen well, it doesn’t feel that hard to build long, meaningful sentences. Subjectively, the difficulty seems to scale roughly proportionally with the length of the poem, rather than quadratically and beyond.
There’s a nice connection here to Sample Space Reducing (SSR) processes. The act of picking letters from a fixed multiset to form words, and removing them as you go, is a SSR. So is sentence formation itself: each committed word constrains the space of acceptable continuations (morphology, syntax, discourse, etc.).
Understanding scaling through history-dependent processes with collapsing sample space, https://arxiv.org/pdf/1407.2775
> Many such stochastic processes, especially those that are associated with complex systems, become more constrained as they unfold, meaning that their sample-space, or their set of possible outcomes, reduces as they age. We demonstrate that these sample-space reducing (SSR) processes necessarily lead to Zipf’s law in the rank distributions of their outcomes.
> We note that SSR processes and nesting are deeply connected to phase-space collapse in statistical physics [21, 30–32], where the number of configurations does not grow exponentially with system size (as in Markovian and ergodic systems), but grows sub-exponentially. Sub-exponential growth can be shown to hold for the phase-space growth of the SSR sequences introduced here. In conclusion we believe that SSR processes provide a new alternative view on the emergence of scaling in many natural, social, and man-made systems.
In my case there are at least two coupled SSRs: (1) the anagrammatic constraint at the line level (letters being consumed), and (2) the layered SSRs of natural language that govern what counts as a well-formed and context-appropriate continuation (from morphology and syntax up through discourse and argumentation). In practice I ended up exploiting this coupling: by reserving or spending strategic words (pronouns, conjunctions, or semantically heavy terms established earlier), I could steer both the unfolding sentence and the remaining letter pool, and explore the anagram space far more effectively than a naive generator.
Very hand-wavy hypothesis: natural language is a complex, multi-layered SSR engine that happens to couple extremely well to other finite SSR constraints. That makes it unusually good at “solving” certain bounded combinatorial puzzles from the inside—up to and including, say, assembling IKEA furniture.
One extra nuance here: in the anagrammatic setting, the coupling between constraints is constitutive rather than merely referential. The same finite multiset of letters simultaneously supports the combinatorial constraint (what strings are formable) and the linguistic constraint (what counts as a syntactically and discursively acceptable move), so every choice is doubly binding. That’s different from cases like following IKEA instructions, where language operates as an external controller that refers to another state space (parts, tools, assembly steps) without sharing its “material” degrees of freedom. This makes the anagram case feel like a toy model where syntax and semantics are not two separate realms but two intertwined SSR layers over one shared substrate—suggesting that what we call “reference” might itself be an emergent pattern in how such nested SSR systems latch onto each other.
Perhaps a controversial view on this particular forum but I find the tendency of a certain type of person* to write about everything in this overly-technical way regardless of whether it is appropriate to the subject matter to be very tiresome ("executing cached heuristics", "constrained the search space").
*I associate it with the asinine contemporary "rationalist" movement (LessWrong et al.) but I'm not making any claims the author is associated with this.
What diction is "appropriate to the subject matter" is a negotiation between author and reader.
I think the author is ok with it being inappropriate for many; it's clearly written for those who enjoy math or CS.
I think it's a trick. It seems to be the article is just a series of ad-hoc assumptions and hypotheses without any support. The language aims to hide this, and makes you think about the language instead of its contents. Which is logically unsound: In a sharp peak, micro optimizations would give you a clearer signal where the optimum lies since the gradient is steeper.
> In a sharp peak, micro optimizations would give you a clearer signal where the optimum lies since the gradient is steeper.
I would refuse to even engage with the piece on this level, since it lends credibility to the idea that the creative process is even remotely related to or analogous to gradient descent.
I wouldn't jump to call it a trick, but I agree, the author sacrificed too much clarity in a try for efficiency.
The author set up an interesting analogy but failed to explore where it breaks down or how all the relationships work in the model.
My inference about the author's meaning was such: In a sharp peak, searching for useful moves is harder because you have fewer acceptable options as you approach the peak.
Fewer absolute or relative? If you scale down your search space... This only makes some kind of sense if your step size is fixed. While I agree with another poster that a reduction of a creative process to gradient descent is not wise, the article also misses the point what makes such a gradient descent hard -- it's not sharp peaks, it's the flat area around them -- and the presence of local minima.
It's a middle school essay that is trying to score points based on the number of metaphors used. Very unappealing and I wouldn't call it technical.
EDIT: For all the people saying the writing is inspired by math/cs, that's not at all true. That's not how technical writing is done. This guy is just a poser.
> I wouldn't call it technical
Fair. Perhaps I should have said it gives the illusion of being technical.
A bit harsh, but I see what you mean. It is tempting to try and fit every description of the world into a rigorous technical straightjacket, perhaps because it feels like you have understood it better?
Maybe it is similar to how scientist get flack for writing in technical jargon instead of 'plain language'. Partly it is a necessity - to be unambiguous - however it is also partly a choice, a way to signal that you are doing Science, not just describing messing about with chemicals or whatever.
I have observed it too, it is heavily inspired by economics and mathematics.
Saying "it's better to complete something imperfect than spend forever polishing" - dull, trite, anyone knows that. Saying "effort is a utility curve function that must be clamped to achieve meta-optimisation" - now that sounds clever
If I was going to be uncharitable, I think there is are corners of the internet where people write straightforward things dressed it up in technical language to launder it as somehow academic and data driven.
And you're right, it does show up in the worse parts of the EA / rationalist community.
(This writing style, at its worst, allows people to say things like "I don't want my tax spent on teaching poor kids to read" but without looking like complete psychopaths - "aggregate outcomes in standardised literacy programmes lag behind individualised tutorials")
That's not what the blog post here is doing, but it is definitely bad language use that is doing more work to obscure ideas than illuminate them
Yes, you articulated my issue in a much better way than I managed to!
I'll be the first to admit I was unable to follow the article because of this.
Just reading the abstract, I have to agree with you.
I mean, I talk like this as well. It's not really intentional. My interests influence the language that I use.
Why is the rationalist movement asinine? I don't know much about it but it seems interesting.
no, we need more of this, the opposite of this is Robin Williams destroying the poetry theory book in dead poeta society, the result was weak kids and one of them commited suicide. More technical stuff in relation to art is a good thing, but its expected that anglosaxon people have allergy to this, they think is somehow socialist or something and they need art to be unfefined etc
I am not sure you watched the same movie I did.
Respectfully, I have no idea what you're talking about. Dead Poets Society is a story and the message of the story isn't that Robin Williams' character is bad.
Are you saying my perspective is anti-socialist? What is "refined" art?
I appreciate this post as I think too many folks focus on the end before understanding what made it there. It's kind of asking what's the movie about before watching it or especially movie trailers that essentially shows way too much.
We should all take some time to better understand what brought us here to be better prepared for general creative work and uniqueness in the future...
"Understanding Poetry, by Dr J. Evans Pritchard, PhD"
lol I cited this exact scene as an example of typical anglosaxon conception of art, now you are crying that art has become shit but any attempt at scientific analysis is taken as a joke when actual poetry is even harder than Code, the amount of data you can compress on a single Word and rhyimes and stuff IS the hardest thing ever, but because you dont want to think someone can do an effort you want the Robin Williams and Dead Poets society to win and make art non scientifically understandable to anyone, if you cant do scientific or technical analysis of art thats your opinión but why the obsession on trashing anyone Who does It?