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.
I trash and restart sketches that took a few minutes to do at most. It's very rare for me to get more than an hour or two in and discard it, I've explored and found a solid foundation long before I put that much work in.
If I was working in a studio environment there's the risk of things like "I spent a couple weeks painting that bg and animating that scene and it was absolutely gorgeous but it got cut when the we took a good hard look at the remaining budget and total runtime and cut the entire sequence it was part of" but that's another matter.
The Draw 50 process is solid. The classical techniques I learnt in training for animation are similar.
In art, while there are mistakes that can really make a drawing "wrong", there is a much wider array of valid end-states. Which is why Generative AI managed to make art way before it managed to write code.
An example of this is Bob Ross' school of incorporating mistakes into the piece. Surely it's something that would fly way less in some types of drawings like anatomical ones, but especially when the motif is fantastical or even abstract, there's fewer possible mistakes to make, so there's less reason to throw something away.
With black ink on paper, quite a few. With oil, less. It is part of the medium: how much is a flash of intuitive certainty and how much is slower planning and execution.
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.
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?
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'm personally a little frustrated with these music theory answers. Trust me folks, these answers are nearly impossible for a non-musician to understand (and even as a musician it's a bit impenetrable).
E minor gives you the exact same amount of options as C major. The options are just shuffled around a little bit. You literally get the same amount of notes in either, just a slightly different set. It isn't any more complex. Listeners aren't going to notice a difference, except one will probably sound happy and one will probably sound sad/angry. The "acceptance volume", to use the blog author's term, isn't any different.
At best, it can change things a little bit for some instruments. For example, with a vocalist, their voice can only go so high. They might be able to hit up to a high C, but not even higher up to an high E. If you're in C major, that's great, the vocalist's highest note (C) is the 'home note' which sounds great (playing a C in C Major makes the song sound like it's 'finished'). If you're E minor, the 'home note' is E, and as mentioned they wouldn't be able to hit that note. So you wouldn't really be able to 'finish' on a high note.
Ultimately, I doubt the author is a musician. It was a strange example to make their point.
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).
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.
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 free...
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...
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.
Changing the words is going to lose some of the low-amplitude frequencies but I'll try.
It's a model for why (call the following X) things get harder when you try to make them more perfect. Let's take X for granted.
You can ask yourself "why is X true?". One model you could have for this is the "finishing touches" model: as a thing gets closer to perfection, identifying imperfections and rectifying them is harder simply out of search constraints (the less of something the harder they are to find).
Another model you could have is the "dimensional model". A thing is great when it's great along many axes. The more dimensions you add, the harder it is to search in them for perfection. Related: the curse of dimensionality.
And here he posits a new model, the "resolution model": the finer the look at what is good, the more 'options' you have at each stage to choose from; it's not just that you make the broad and then refine within, but that you are actually building the refined thing from the beginning.
He then tries to show how some kinds of creation tolerate movement in the parameter space better.
No model is perfect, so each of these ideas captures some attribute of the difficulty and maps it to a mental structure that is more easily manipulable by the modeler.
The typical owl drawing is a few circles and then the more owly bits, and then the feathers on the owly bits, and then the shadows on the feathers on the owly bits. And this is a method to reduce the kind of problem he's talking about. But if you want to make the perfect owl, perhaps there is an element of making your circle just so, already accounting for the shadows on the feathers on the owly bits before any of the precursors are made.
Anyway, this is imperfect because I am necessarily shaving off the hair on the ball to show you it's spherical. And his entire model is that the hair determines the ball.
I had a cabinet built and the guy doing the work pointed out that the human eye is really great at detecting line-line deviation; but, building to & correcting for that deviation requires working across the whole surface. He was making an area-effort to linear-quality argument. He said every time you halved the gap, it quadrupled the effort. Also, he said that was what saw-dust & glue were for.
Bands and artists do sometimes record a banger very quickly, and no, the effort is not in muscle memory or practice.
If it was in muscle memory it would be repeatable feat, and it really isn't.
Some work is technically polished and you can see/hear the effort that went into it.
But there's a point where extra effort makes the work less good. Music starts to sound stale and overproduced, art loses some of its directness, writing feels self-conscious.
Whatever the correlation between perceived artistic merit and effort, it's a lot more complex than this article suggests.
"But there's a point where extra effort makes the work less good."
This happens frequently in mixing and mastering audio tracks. You pile up incremental changes that all seemed good at the time. Then you go back and listen to a recording from 20 revisions ago and it sounds better than your current "best" effort.
I just saw Herb Alpert and the Tijuana Brass in concert last night (he's 90 and can still play that horn!) and one of the things he talked about was recording his vocals for This Guy's In Love With You written by Burt Bacharach and Hal David. He's a great musician but not much of a trained singer, and he said when he got into the studio, he sat down and did a relaxed first practice take, expecting he'd go and record one "for real".
Instead, they told him his first take was absolutely perfect - and it went on to be Bacharach/David's first #1 hit song.
Sometimes you get lightning in a bottle, and you just don't mess with it because it's perfect even in its imperfections.
And sometimes for some types of music they will intentionally re-do takes until the singer is a little hoarse and irritated to get a more 'raw' performance.
I discussed this premise with my LLM and we came to this following conclusion which I find quite elegant:
> In any bounded system under feedback, refinement produces diminishing returns and narrowing tolerance, governed by a superlinear precision cost.
> There isn’t one official name, but what you’ve articulated is essentially a unified formulation of the diminishing-returns / sensitivity-amplification law of creation — a pattern deep enough that it keeps being rediscovered in every domain that pushes against the limits of order.
The more mileage you get, the easier it is to see the mistakes in your old art (if you’re improving lol)
The more refined your technique is, the harder it will be to discern mistakes and aesthetic failures.
Eventually you might come to a point where you can’t improve because you literally don’t see any issues. That might be the high water mark of your ability.
Isn't there simply a human tendency to try and find a magic formula in what is simply survivor bias?
As creative projects (software, painting...) we finish or satisfactorily achieve relatively few items. And except for the most repetitive of us, these achievements are pretty different. Wide space, chaos, few satisfactory products by comparison. That doesn't bode well for "magic recipe". All the way to "rules of thumb" that we take fun in violating.
So there are two issues in there: We have more taste than skill and so many of our attempts will disappoint us no matter what. And we will obsess on trying to find a magic formula - when it's rather likely that there isn't one. The "space" is large and chaotic and we might want to reassure ourselves instead that serendipity has something to do with it.
Is there then place for rules of thumb? Whatever let's us get to work in the morning, I guess. For me, I do like the recognition of past track record: with a bit of age hindsight, I know I can do it - no need to dispair. That is useful and reassuring. If I just try some more - in ever varying manners, it will happen.
One place for "rules of thumbs" is in them being tropes. We can get some impact on the viewer by violating them. There is a trope of learning the rules so you can violate them. For example a Rule of Thirds - can be fertile grounds for getting at the viewer. The rule doesn't do much for us, and we have no problem violating it, but our less savvy viewers might remember it and get one more whiff of meaning from the violation. And if we are less concerned with our own satisfaction and more interested in sales, we might pay attention to "what's popular these days" and produce some of that. Not all artists are dead set on personal achievement at the cost of sales. A slightly different look at such rules.
After this you get to the commercial market for creative works. In a market that with a lot of options for consumers that are equally accessible, similarly priced, and infinitely replicable, power-law-curve nonlinear returns on quality is the norm.
So, there is huge motivation to put in “just a bit more effort”.
Author here. Somehow the worst thing I ever wrote is on the front page of HN.
I wrote this fast so there's jargon and bad prose.
The title is deliberately dry and bland so I wasn't expecting anyone to click it. Also I slightly changed my mind on some of the claims .. might write up later.
The main reason I like to think of creative work in a more abstract/formal/geometric way (acceptance volume, latency, sampling) is it's easier for me categorize tasks, modalities and domains and know how to design or work around it. It's very much biased by more own experiences making things.
Also, abstract technical concept often come with nice guarantees/properties/utils to build on .. some would say that's their raison d'être.
Re comments:
* "this is just diminishing returns" -- ok and this is a framework for why: the non-worsening region collapses, so most micro-edits fail
* "bands record bangers in an hour" –– practice tax was prepaid. The recording session is exploitation/search riding on cached heuristics imo (and it still takes hours of repeated recording/mixing/producing to actually produce a single album track).
* music key example –– yes I should've picked a different one. Main point was that some choices create wider tolerance (arrangement/range/timbre) even if keys are symmetric in equal temperament
> Drawing takes forever because you're exploring AND refining simultaneously.
> We don't "rehearse" a specific drawing, we solve a novel problem in real-time. There's no cached motor sequence to execute.
When you have been drawing long enough there are a lot of cached motor sequences to execute and modify. A lot of art training is simply filling this cache: spend a few hours every week drawing the human body from different angles, in a year or three you'll be able to make it up from pretty much any angle. Add in another twenty years of doing that and experimenting ways to make your tools do more of the work for you and you can dash off "sketches" that a beginner would consider finished paintings that took days to do.
The picture of the solution space in 3D makes a great point - we see a narrow hill that leads to a global maximum (i.e. a great result) in a solutions search space that otherwise has a very obvious & wide hill that produces "okay" results. Going from the okay & safe results to a great result means taking the risk of going back down the hill of shittier solutions.
He points out that generative AI will tend to produce results that land on that big wide hill. It's the safe hill, and has the most results. This is perhaps where taste (as a proxy of experience) trumps AI.
Interesting to tie this to the finishing stage of any work. I was definitely thinking about software development in that situation. I would argue it's similar to drawing as he mentions in the FAQ - we're solving a novel problem, as we start implementing a solution we might discover it is inappropriate and have to change to a different part of the solution space.
34 comments
[ 3.7 ms ] story [ 54.6 ms ] threadThere'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.
I trash and restart sketches that took a few minutes to do at most. It's very rare for me to get more than an hour or two in and discard it, I've explored and found a solid foundation long before I put that much work in.
If I was working in a studio environment there's the risk of things like "I spent a couple weeks painting that bg and animating that scene and it was absolutely gorgeous but it got cut when the we took a good hard look at the remaining budget and total runtime and cut the entire sequence it was part of" but that's another matter.
The Draw 50 process is solid. The classical techniques I learnt in training for animation are similar.
An example of this is Bob Ross' school of incorporating mistakes into the piece. Surely it's something that would fly way less in some types of drawings like anatomical ones, but especially when the motif is fantastical or even abstract, there's fewer possible mistakes to make, so there's less reason to throw something away.
*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.
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)?
E minor gives you the exact same amount of options as C major. The options are just shuffled around a little bit. You literally get the same amount of notes in either, just a slightly different set. It isn't any more complex. Listeners aren't going to notice a difference, except one will probably sound happy and one will probably sound sad/angry. The "acceptance volume", to use the blog author's term, isn't any different.
At best, it can change things a little bit for some instruments. For example, with a vocalist, their voice can only go so high. They might be able to hit up to a high C, but not even higher up to an high E. If you're in C major, that's great, the vocalist's highest note (C) is the 'home note' which sounds great (playing a C in C Major makes the song sound like it's 'finished'). If you're E minor, the 'home note' is E, and as mentioned they wouldn't be able to hit that note. So you wouldn't really be able to 'finish' on a high note.
Ultimately, I doubt the author is a musician. It was a strange example to make their point.
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).
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.
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 free...
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...
> 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.
Changing the words is going to lose some of the low-amplitude frequencies but I'll try.
It's a model for why (call the following X) things get harder when you try to make them more perfect. Let's take X for granted.
You can ask yourself "why is X true?". One model you could have for this is the "finishing touches" model: as a thing gets closer to perfection, identifying imperfections and rectifying them is harder simply out of search constraints (the less of something the harder they are to find).
Another model you could have is the "dimensional model". A thing is great when it's great along many axes. The more dimensions you add, the harder it is to search in them for perfection. Related: the curse of dimensionality.
And here he posits a new model, the "resolution model": the finer the look at what is good, the more 'options' you have at each stage to choose from; it's not just that you make the broad and then refine within, but that you are actually building the refined thing from the beginning.
He then tries to show how some kinds of creation tolerate movement in the parameter space better.
No model is perfect, so each of these ideas captures some attribute of the difficulty and maps it to a mental structure that is more easily manipulable by the modeler.
The typical owl drawing is a few circles and then the more owly bits, and then the feathers on the owly bits, and then the shadows on the feathers on the owly bits. And this is a method to reduce the kind of problem he's talking about. But if you want to make the perfect owl, perhaps there is an element of making your circle just so, already accounting for the shadows on the feathers on the owly bits before any of the precursors are made.
Anyway, this is imperfect because I am necessarily shaving off the hair on the ball to show you it's spherical. And his entire model is that the hair determines the ball.
If it was in muscle memory it would be repeatable feat, and it really isn't.
Some work is technically polished and you can see/hear the effort that went into it.
But there's a point where extra effort makes the work less good. Music starts to sound stale and overproduced, art loses some of its directness, writing feels self-conscious.
Whatever the correlation between perceived artistic merit and effort, it's a lot more complex than this article suggests.
This happens frequently in mixing and mastering audio tracks. You pile up incremental changes that all seemed good at the time. Then you go back and listen to a recording from 20 revisions ago and it sounds better than your current "best" effort.
Instead, they told him his first take was absolutely perfect - and it went on to be Bacharach/David's first #1 hit song.
Sometimes you get lightning in a bottle, and you just don't mess with it because it's perfect even in its imperfections.
> In any bounded system under feedback, refinement produces diminishing returns and narrowing tolerance, governed by a superlinear precision cost.
> There isn’t one official name, but what you’ve articulated is essentially a unified formulation of the diminishing-returns / sensitivity-amplification law of creation — a pattern deep enough that it keeps being rediscovered in every domain that pushes against the limits of order.
The more refined your technique is, the harder it will be to discern mistakes and aesthetic failures.
Eventually you might come to a point where you can’t improve because you literally don’t see any issues. That might be the high water mark of your ability.
As creative projects (software, painting...) we finish or satisfactorily achieve relatively few items. And except for the most repetitive of us, these achievements are pretty different. Wide space, chaos, few satisfactory products by comparison. That doesn't bode well for "magic recipe". All the way to "rules of thumb" that we take fun in violating.
So there are two issues in there: We have more taste than skill and so many of our attempts will disappoint us no matter what. And we will obsess on trying to find a magic formula - when it's rather likely that there isn't one. The "space" is large and chaotic and we might want to reassure ourselves instead that serendipity has something to do with it.
Is there then place for rules of thumb? Whatever let's us get to work in the morning, I guess. For me, I do like the recognition of past track record: with a bit of age hindsight, I know I can do it - no need to dispair. That is useful and reassuring. If I just try some more - in ever varying manners, it will happen.
One place for "rules of thumbs" is in them being tropes. We can get some impact on the viewer by violating them. There is a trope of learning the rules so you can violate them. For example a Rule of Thirds - can be fertile grounds for getting at the viewer. The rule doesn't do much for us, and we have no problem violating it, but our less savvy viewers might remember it and get one more whiff of meaning from the violation. And if we are less concerned with our own satisfaction and more interested in sales, we might pay attention to "what's popular these days" and produce some of that. Not all artists are dead set on personal achievement at the cost of sales. A slightly different look at such rules.
So, there is huge motivation to put in “just a bit more effort”.
And, thus you get Crunch Time in gamedev!
90% of the project takes 90% of the time and the other 10% of the project takes another 90% of the time.
I wrote this fast so there's jargon and bad prose. The title is deliberately dry and bland so I wasn't expecting anyone to click it. Also I slightly changed my mind on some of the claims .. might write up later.
The main reason I like to think of creative work in a more abstract/formal/geometric way (acceptance volume, latency, sampling) is it's easier for me categorize tasks, modalities and domains and know how to design or work around it. It's very much biased by more own experiences making things.
Also, abstract technical concept often come with nice guarantees/properties/utils to build on .. some would say that's their raison d'être.
Re comments: * "this is just diminishing returns" -- ok and this is a framework for why: the non-worsening region collapses, so most micro-edits fail
* "bands record bangers in an hour" –– practice tax was prepaid. The recording session is exploitation/search riding on cached heuristics imo (and it still takes hours of repeated recording/mixing/producing to actually produce a single album track).
* music key example –– yes I should've picked a different one. Main point was that some choices create wider tolerance (arrangement/range/timbre) even if keys are symmetric in equal temperament
> We don't "rehearse" a specific drawing, we solve a novel problem in real-time. There's no cached motor sequence to execute.
When you have been drawing long enough there are a lot of cached motor sequences to execute and modify. A lot of art training is simply filling this cache: spend a few hours every week drawing the human body from different angles, in a year or three you'll be able to make it up from pretty much any angle. Add in another twenty years of doing that and experimenting ways to make your tools do more of the work for you and you can dash off "sketches" that a beginner would consider finished paintings that took days to do.
The picture of the solution space in 3D makes a great point - we see a narrow hill that leads to a global maximum (i.e. a great result) in a solutions search space that otherwise has a very obvious & wide hill that produces "okay" results. Going from the okay & safe results to a great result means taking the risk of going back down the hill of shittier solutions.
He points out that generative AI will tend to produce results that land on that big wide hill. It's the safe hill, and has the most results. This is perhaps where taste (as a proxy of experience) trumps AI.
Interesting to tie this to the finishing stage of any work. I was definitely thinking about software development in that situation. I would argue it's similar to drawing as he mentions in the FAQ - we're solving a novel problem, as we start implementing a solution we might discover it is inappropriate and have to change to a different part of the solution space.