To paraphrase that quote about hydrogen: Give gradient descent a few billion parameters, and it starts wondering where it came from and what does it all mean.
Constraints forces creativity. Some well chosen constraints are aesthetics rules that helps you land pleasing results. Poetry has a long history on that matter.
Another example of constrained creativity is early to mid nineties electronic music.
> Inspired by the Twitter logo, which is made from 13 perfect circles
Compared to that, the new logo doesn't have a circle (segment) anywhere to be seen (unless you consider straight lines as circle segments with the center located at infinity of course), and is simply the "mathematical double-struck capital X" from an unknown but probably pre-existing font (apparently Monotype's "Special Alphabets 4" comes close, but isn't identical, according to https://tweethunter.io/blog/how-to-write-twitter-x-iphone-ma...).
Im curious what the process looks like to implement this. It seems like it would be easiest to start with the animal using only perfectly(?) curved lines and then complete them into circles after the fact. Although that seems kind of pointless and I imagine they start with circles. And I guess it would hard to have a curve from a perfect circle without the circle?
I just have a hard time imagining you start with circles, lay them down (resize as needed) and continue. I mean I guess that doesnt sound so crazy after I say it... it just seems like it would add a lot of extra noise to the image that would make it much harder to draw.
I can't speak to this but I took a drawing class a long time ago. I'm not very good but it was a lot of drawing circles. When you see people freehand stuff it's kind of wild but that's not how people learn how to draw they're just very good at it from practice. Most of learning is drawing very basic shapes, usually circles, and erasing parts that don't make sense and continuing.
> drawing very basic shapes, usually circles, and erasing parts that don't make sense
There's a hilarious Spongebob bit [0] where Squidward is teaching an art class, and he starts off in that exact manner of trying to draw a perfect circle, only to have Spongebob subvert the entire idea. The whole episode is artistic gold IMO.
I have been practicing art a lot lately. You can draw just about anything using spheres, cubes, cylinders, and cones. You start off with the 2d versions.
I stopped after a few classes but I was amazed at how good I got in a short amount of time after learning how to break stuff down which isn't something I really thought about before. By all metrics I'm still a pretty terrible drawer but prior to that stick figures would have been challenging.
Another good resource for learning how to draw realistically is the book: "Drawing on the Right Side of the Brain". The premise is that your brain wants to take shortcuts and group/chunk things together on what they should look like, instead of what things actually look like. But even a rectangle in real life has non-right-angles because of perspective, etc.. And if you draw what you actually see, then the drawings come out correct. Some of the exercises are copying other drawings placed upside-down, so that you brain doesn't try to over-interpret things. I can't recommend this enough if you want to go from a beginner to something respectable in drawing abilities.
I read the book and loved it (about 15 years ago). There’s no royal road to becoming an artist but lots of joy along the way. Whatever the path, enjoy it!
Drawing from circles, squares, triangles, etc. in art is called "construction" and is definitely a foundational technique. It really is amazing how much easier drawing becomes once it's understood (and practiced).
"While sketching, I kept track of the number of circles I was using, counting one for every curve."
After sketching an animal, it should be easier to adjust the image by inserting/removing/moving circles.
I suppose the thing the circle is really informing is the "perfectness" of the curve. You cant just draw in curves and extend it to a circle (wont be perfect). I guess Im not sure how you get "perfect" curves.
I suspect its a stencil or something. So in some sense the circle does exist first, even if they only draw the curve from it initially (before marking it up with the full circle after the fact).
If I were trying to do something like this I would sketch it out first with imperfect curves and then worry about making it perfect once I was at the computer. It would look slightly different but I don’t think it would make that much of an impact in the initial design process.
What's wild is how much clarity and personality you can get from that process. Instead of adding noise, it forces simplification, which actually helps with visual clarity
He was done a great deal of injustice when he was passed over for the Turing award that was given to Hinton, Bengio, LeCun.
Then there is this from his blog --
Dec 2024: Sadly, the Nobel Prize in Physics 2024 for Hopfield & Hinton is a Nobel Prize for plagiarism. They republished methodologies developed in Ukraine and Japan by Ivakhnenko and Amari in the 1960s & 1970s, as well as other techniques, without citing the original papers. Even in later surveys, they didn't credit the original inventors (thus turning what may have been unintentional plagiarism into a deliberate form). None of the important algorithms for modern Artificial Intelligence were created by Hopfield & Hinton. Details in the recent technical report, with lots of references, links, and facts.
I'm not sure whether or not he did this first, but it's very similar to an extremely impressive, but old and well-known illustration of the power of Fourier analysis in which you construct a "Fourier epicycle" (think: machine made of circular gears of different ratios) that can sketch any image. 3blue1brown has a great video on Fourier Epicycles but you can also get the idea here https://mathematica.stackexchange.com/questions/171755/how-c...
Also of potential interest is Kempe's Universailty Theroem which states you can draw any (polynomial) shape with a set of mechanical linkages. Like one that will sign your name.
This is so true. I took a figure drawing class in college and we were instructed to draw with a cloth and charcoal dust. Easily some of the best work I made came out of that restriction.
Architecture too. The worst building come from architects given a blank page to start with. Constraints, and sympathy for the surrounding built environment produce great work.
I miss being creative, before I knew how to make front end UIs I had crazy ideas but then became grounded. This one isn't super crazy but I like those vertical buildings.
Tangent, with a dark/colorful theme in an editor the minimap looks like a city scape
My aunt grifted me “Animali Compassati” when I was a kid… A small book with instructions for animals that you could draw with a compass.
The site is unclear somehow… but the instructions were pretty great in the book.
This page feels like an AI traveled back in time and (faked) the date.
[edit] Nevermind. I'm being too harsh. The creator was obviously having fun and being creative. That's cool. I think if nothing else this just proves how jaded and skeptical about clever artwork I've become in the past few years.
If you pay closer attention, you can see that some of the designs rely on very deliberate placements of circles; for example, eyes of the monkey and owl, and the nose of the whale.
Those are just the obvious ones that I can immediately spot — there was probably a lot of careful consideration into the placement of circles in order to facilitate good looking arcs and circles that bring the animals to "life".
Just doing a Google search for "animals made from circles", you get the usual header full of "Images" and "Videos" crap, then in the actual results links, you have the usual Pinterest linkslop, Facebook linkslop, Reddit linkslop, a bunch of articles written by the designer (now we're getting somewhere). OP's link is finally on page 4 of the search results.
For me, searching "animals made from circles", your comment put this HN thread as the #1 result while the #2 result was a syndicated article about the linked post. When I get more specific and search "animals drawn only from circles" it turns up the linked post as the first result. But my results may be more specific partly because I don't use ad blockers.
I was thinking that this would be low-hanging fruit for a model. The parameter space is so tiny compared to what a diffusion model already has to deal with...
More generally, any animal that cannot be drawn in 12 circles cannot be drawn in 13 circles when riding a bicycle. By recursion, no animal can be drawn when riding a stack of seven bicycles.
You also need to encode the painted areas somehow. They are not only intersections on K shapes, but sometimes exclusions as well (like (A^B)/C). Two ways come to mind:
(1) Listing closed curves by vertices. Each vertex of a painted area is an intersection of two or more circles, and delimits a section of a circle. So the section of circles that enclose a circle can be encoded each by the union of:
(1.1) A circle (index);
(1.2) A 2nd circle (index) that intersects the 1st on a first point;
(1.3) A bit identifying the (first) intersection (because there may be 2 possible);
(1.4) A 3rd circle (index) that intersects the 1st on a second point;
(1.5) A bit identifying the (second) intersection.
Note the base circle would be the first intersection of a subsequent section of this closed curve, and the 3rd circle would be the subsequent base circle. So 1/2/3 won't be necessary for subsequent curves. So only (K+2) indices + (K+1) bits are necessary for this encoding.
Total ~K log2(K)+K bits. I hypothesize (left to the reader :)) a closed curve should contain at most 2x13 points. There can be at most 2^13 distinct regions however, so each figure (Animal) can be encoded with less than that many curves per figure. So each figure (Animal) can be encoded with less than 2^13 x 26x(5+1) bits =~ 1.3Mbit.
But that's mostly pathological cases, if each Animal must be a fully connected area, then that might reduce (hypothesis above) to at most only
26x(5+1) bits = 156 bits, or 20 bytes!
I left out a problem which area shapes encoded within each other (like eyes). In that case you need at most another 156 bits per inner cutout shape.
(2) Alternatively, you could use boolean operations to encode each shape. Also left as a fun problem :)
I tend to wonder if stuff like this is an informative boundary on AI capabilities. I mean, you can't ask a LLM today to do that (AFAICT). "Here's a simply-specified but extremely broad search space, solve this problem in it" isn't something that fits the model. But it's a relatively common (if not "easy") task human beings like to show off.
What needs to change to enable this kind of exploration?
Nope, but this post is such a neat illustration of the richness of "life" that fits into 39 real parameters (each circle can be coordinatized as 3 real numbers: one for its radius and two for its center) that my first thought on seeing it was also "no surprise then that a matrix with a million entries can talk like an erudite person".
Wouldn't you also need a two parameters for the arc starting position and stopping position for each circle, and then a few more to identify the areas that need to be filled, along with the color?
It should be obvious that this is entirely up to the reader. Take some responsibility for your own happiness. Nobody else is responsible for your enjoyment of anything.
Well, sort of? I mean, I've seen plenty of clever art in my life. I'm still figuring out AI. I posted that in the hope that someone in the community here would show up with something insightful to say.
Yes, it is impossible. People will think about things they find interesting regardless of your (dis-)approval. Who are you even calling "obsessive?" The collective of people who dare to mention algorithms you don't like? I mean, what the fuck?
Actually, the (in)famous "sparks of general intelligence paper" about GPT-4 included tasks such as "Draw a unicorn in TikZ" which really is not that far off from this task. There were also examples for drawing cars/trucks/cats etc with SVG.
But I do think that evolutionary algorithms or MCMC variants could do a better job of this, especially if paired with an auxiliary model for scoring their intermediate results.
Yes, this has been done in many forms with other algorithms. You score each generation with a model like CLIP, for example, and then you can evolve 'Mona Lisa made of triangles', say. A constraint like 'exactly 13 circles' will work fine. (And you might experiment with loosening it, like generating a lot of candidates with 5-30 circles each, as a 'library' or 'seeds', before shrinking them all towards 13, to see if you get novel animal designs which are find to find if you simply start the obvious way with 13 circles initialized to random points & sizes.)
I was thinking it could, actually, given a feedback loop. The tool use would a json that takes 13 circles, each with x, y position, radius, and whether it's filled in or empty, and output an image. It could look at the image and iterate.
With four parameters I can fit an elephant, and with five I can make him wiggle his trunk.
https://en.m.wikipedia.org/wiki/Von_Neumann%27s_elephant
I did a few back in my day as a designer:
1. https://dribbble.com/shots/1909369-Liberty-Eagle-Arms
2. https://dribbble.com/shots/1553151-Flint-mark-icons
That first one is some of my best work.
Another example of constrained creativity is early to mid nineties electronic music.
Compared to that, the new logo doesn't have a circle (segment) anywhere to be seen (unless you consider straight lines as circle segments with the center located at infinity of course), and is simply the "mathematical double-struck capital X" from an unknown but probably pre-existing font (apparently Monotype's "Special Alphabets 4" comes close, but isn't identical, according to https://tweethunter.io/blog/how-to-write-twitter-x-iphone-ma...).
I just have a hard time imagining you start with circles, lay them down (resize as needed) and continue. I mean I guess that doesnt sound so crazy after I say it... it just seems like it would add a lot of extra noise to the image that would make it much harder to draw.
There's a hilarious Spongebob bit [0] where Squidward is teaching an art class, and he starts off in that exact manner of trying to draw a perfect circle, only to have Spongebob subvert the entire idea. The whole episode is artistic gold IMO.
0: https://www.youtube.com/watch?v=vTlpFEvmxdM
https://www.reddit.com/r/restofthefuckingowl/comments/6f71jm...
https://imgur.com/how-i-feel-when-somebody-gives-me-advice-g...
https://www.amazon.com/Drawing-Right-Side-Brain-Definitive/d...
https://kk.org/cooltools/drawing-on-the-right-side-of-the-br...
"While sketching, I kept track of the number of circles I was using, counting one for every curve." After sketching an animal, it should be easier to adjust the image by inserting/removing/moving circles.
I suspect its a stencil or something. So in some sense the circle does exist first, even if they only draw the curve from it initially (before marking it up with the full circle after the fact).
https://www.reddit.com/r/interestingasfuck/comments/txdimd/t...
[1] https://en.wikipedia.org/wiki/Low-complexity_art
My son is a big fan of bytebeat [2], which is also a low complexity art, but music.
[2] https://dollchan.net/bytebeat/#4AAAA+kUryC/X0CixswNhQyM1Q01N...
Then there is this from his blog --
Dec 2024: Sadly, the Nobel Prize in Physics 2024 for Hopfield & Hinton is a Nobel Prize for plagiarism. They republished methodologies developed in Ukraine and Japan by Ivakhnenko and Amari in the 1960s & 1970s, as well as other techniques, without citing the original papers. Even in later surveys, they didn't credit the original inventors (thus turning what may have been unintentional plagiarism into a deliberate form). None of the important algorithms for modern Artificial Intelligence were created by Hopfield & Hinton. Details in the recent technical report, with lots of references, links, and facts.
https://people.idsia.ch/~juergen/physics-nobel-2024-plagiari...
Also, AlphaFold is great but hardly an innovation. David Baker deserved it 100%.
https://academic.oup.com/plms/article/s1-7/1/213/1570315?log...
http://www.koutschan.de/data/link/
Often depicting slices of vegetables, animals..
From few circles to hundreds
So if we remove the condition of 13, everything is in fact made of circles only!
Tangent, with a dark/colorful theme in an editor the minimap looks like a city scape
https://www.danielenannini.it/en/portfolio/animali-compassat...
https://www.compasses-zoo.net/compasses-zoo/index.php
[edit] Nevermind. I'm being too harsh. The creator was obviously having fun and being creative. That's cool. I think if nothing else this just proves how jaded and skeptical about clever artwork I've become in the past few years.
Those are just the obvious ones that I can immediately spot — there was probably a lot of careful consideration into the placement of circles in order to facilitate good looking arcs and circles that bring the animals to "life".
This type of content is becoming rarer on the internet nowadays.
What animals cannot be accurately depicted with 13 circles?
A mature house centipede has 15 pairs of legs. You can probably get the point across with a portion of that, and use two parts of a circle for 2 legs.
But not that artsy as the OG.
Circle circles_[13];
}
(1) Listing closed curves by vertices. Each vertex of a painted area is an intersection of two or more circles, and delimits a section of a circle. So the section of circles that enclose a circle can be encoded each by the union of:
(1.1) A circle (index); (1.2) A 2nd circle (index) that intersects the 1st on a first point; (1.3) A bit identifying the (first) intersection (because there may be 2 possible); (1.4) A 3rd circle (index) that intersects the 1st on a second point; (1.5) A bit identifying the (second) intersection.
Note the base circle would be the first intersection of a subsequent section of this closed curve, and the 3rd circle would be the subsequent base circle. So 1/2/3 won't be necessary for subsequent curves. So only (K+2) indices + (K+1) bits are necessary for this encoding.
Total ~K log2(K)+K bits. I hypothesize (left to the reader :)) a closed curve should contain at most 2x13 points. There can be at most 2^13 distinct regions however, so each figure (Animal) can be encoded with less than that many curves per figure. So each figure (Animal) can be encoded with less than 2^13 x 26x(5+1) bits =~ 1.3Mbit.
But that's mostly pathological cases, if each Animal must be a fully connected area, then that might reduce (hypothesis above) to at most only 26x(5+1) bits = 156 bits, or 20 bytes!
I left out a problem which area shapes encoded within each other (like eyes). In that case you need at most another 156 bits per inner cutout shape.
(2) Alternatively, you could use boolean operations to encode each shape. Also left as a fun problem :)
The Procedural Animation https://www.youtube.com/watch?v=qlfh_rv6khY
Gibbon: Beyond the Trees - Wolfire Games https://www.youtube.com/watch?v=KCKdGlpsdlo
What needs to change to enable this kind of exploration?
Then if you want to factor out rotations, drop another parameter and say the 1st explicit circle lies on the x axis.
Arcs are just intersection of circles, so they are implicit, as far as I can tell.
But I do think that evolutionary algorithms or MCMC variants could do a better job of this, especially if paired with an auxiliary model for scoring their intermediate results.