you joke about toggling Caps Lock to write capital letters every time, but that's actually how my history teacher used to write capital letters, and she got pretty efficient at it
I had a roommate in college that did the same thing. Didn't matter how much you told him about the shift key, he'd still do the caps lock thing every time.
@@Damaniel3 my dad still does that, its a thing for all those "hunt and peck" typers that never really learned how to type correctly or with more than 1 finger :D
Numbers are already uppercase. When they were making telegrams, Morse, teletypes, punch cards and so on, it's always uppercase letters and numbers first and lowercase later added on (this is why ASCII is roughly in the order of control codes, numbers, uppercase, lowercase, with punctuation sprinkled in between) Typographically, uppercase has primary features going from the baseline to the ascender, and no features down to the descender, while lowercase letters have primary features only up to the midline, and only minor features like stems up to the ascender and down to the descender. Numbers are all primarily from the baseline to the ascent, and thus are uppercase. Thank you for coming to my TED talk.
I work in software... i dream of one day being able to just make 5+ different user interfaces for some random project... frankly, this guy really is just on a whole other level.
I imagine you could make a program to first take the input letter and generate a chessboard based on it, then have each chess algorithm play against itself from that position for a set number of turns and use that new board state as the output character. Actually designing or training the algorithm to create a board that leads to interesting results for each chess algorithm would be the hard part.
There needs to be a word for "Useless but Entertaining Experimental Computer Stuff" because that is one of my greatest joys in life. It's also just obscure enough to be kind of hard to find using conventional youtube search methods, and now I've found a channel that does some very excellent UbEECS and great art to boot. Absolutely earned a subscription.
Just goes to show how arbitrary and re-interpretable games can be. You don't _have_ to maximise efficiency, or use out-of-the-box tactics, or even aim to win; if the true goal of Chess this entire time was just to make fun shapes, we'd never know about any of that, which is a poetic window into an alternate objective.
"You will fail because you merely aim to win, while I aim to make beautiful shapes." - A chess agorithm that optimizes for the board looking like the letter 'C', moments before giving up both of its bishops.
would've been better if it had done a game tree traversal to find the one most likely to create a C, currently it's easy for it to make a terrible move where the opponent can immediately ruin the C
@@alex15095 Now I'm imagining the perfect "chess" algorithm that will only ever goad its opponent into making a C out of all the pieces, and if its opponent ruins the C, the algorithm just keeps pushing the board until it can only ever look like a C.
I just found out you're a contributor at DFX. I've used and loved your plugins for a few years now. You're both incredibly, and intimidatingly talented! Thank you!
Cool! Thanks and thanks for using those! Yes, actually we have been working on destroy fx for the first time in many years... some new releases soon :)
It would be really cool to try this with Chinese characters. It would be easiest to do it with just Traditional and Simplified characters (for example, "horse" was simplified from 馬 to 马). However, all of the characters have been evolving even before that, and many can be traced all the way back to the bronze age; a few thousand can even be traced back on 3,000 year old oracle bones. And those kind of resemble stone-age pictographs. So maybe you could come up with a program that takes any doodle of an object, and puts it through thousands of years of evolution to instantly make a modern character. And then maybe see what it would look like however many years in the future.
Kana would be interesting too! It could be trained to “hiraganaize” and “katakanaize”, and then let loose. You could do wild stuff, like ask it for the katakana version of the letter D. The letter case data could also be used to create a lowercase を. Kanji/hanzi could also be added into the mix too, cause I’m very interested what a capital 議 would look like.
The English alphabet is kinda the best case (ha) scenario for this, it has simple shapes and they're all very distinct, spatially speaking, from each other. Chinese (and similar languages) have extremely complex shapes (even Simplified Chinese) that have no significant spatial relationships. It would require a stupidly gigantic network and a massive amount of training time.
Pushing the saying "because we can" to the next level! These videos are so entertaining and somehow reassuring me in my fear of wasting my life...why? I don't know.
These videos lull me into a false sense that I am actually engaging in worthwhile thought, and it is moments like the one you described that snap me back to reality and make me realize that I am fully down the rabbit hole of nonsense.
19:27 this is why you use the softmax function on your categorical outputs. It doesn't just constrain the outputs to be between 0 and 1, it also makes them sum to 1 (and does it in a differentiable way, so you can backprop across it). With softmaxed probabilities, maximizing the 'f' score is actually giving you the thing that is "most like an f while also being least like everything else".
makes me think the model could perform better with some deep learning tinkering - since he says "the layer is initialized to a diagonal" it also seems that he's using a dense NN instead of a CNN, for instance
@Wave I think he said it initialized to the identity matrix. Which confused me as well.. should be using random weights. And yeah a model which uses convolutions and deconvolutions for an encoder and decoder would probably be superior if he is not doing that. Although with input and output sizes of 36x36 it's probably not *that bad* of an approach for just screwing around. The model will definitely take longer to train and probably be significantly less efficient, but if you throw enough parameters at it, it would probably compensate somewhat for the bad architecture.. and that's all machine learning really is right? Why spend time using my brain when I can just upgrade my gpu?
@@derpnerpwerp the reason he initialized it to the identity matrix is because it was added during training and randomizing the weights would make all prior training useless
I mean this entirely positive, this video is somehow defeating my insomnia within the first 5 minutes. Organized, rhythmic, intriguing. Absolutely fantastic, I hope I fall asleep before finishing it so I can rewatch it tomorrow night. :)
when you said logical conclusion, I was sort of expecting upper and lower case numbers? regardless, this is my favorite time of year and you can do no wrong. I admire everything you create and I can't wait for whatever comes next!
It Could also be called Supercaps, meaning you could have a supercaps lock button, or a Megashift button for your keyboard. But since you have lowestcase, you need to invent 2 other keys, called undershift and Lowcaps lock. I don’t know if a keyboard inventor will invent this, but it would be awesome.
I think using mathematical double-struck letters (The ones used for sets) would be a very good standard for Uppestcase because they're easily recognizable and at the same time are well set apart.
Every time you upload, I think to myself, "wait, who?" Then I watch, and it's great, and then go back through your old videos, and remember why I turned on notifications in the first place. Always a pleasure.
I aspire to be the kind of man that could make people feel the way the end of every tom7 video has me feeling, just genuine happiness and amazement at what I just witnessed.
"it doesn't care about chess, it just cares about making Cs" Dude, I joined a chess club in high school for the extracurricular requirement. I know exactly how that AI feels.
2:14 Why do easy things in 90 minutes when you can spend weeks automating something inefficient that is also needlessly complicated? This is the stuff I live and breath for!
I love how you take an odd machine learning problem and then go through a bunch of odd but genius solutions you come up with. Please make more of these types of videos.
As always, this was glorious, and the ending was just...so good. I'm glad to see you back making things, and I'm very excited to see what madness you do next!
You should've included other languages upper and lowercase letters in the training data, there's no possible way that would contribute to more complexity or confusion
It surely won't get confused with these :) - Nu (ν,Ν) looks like V and N - Eta (η,Η) looks like N and H - Gamma (γ,Γ) looks like Y and R - Upsilon (υ,Υ) looks like U and Y - Mu (μ,Μ) looks like U and M - Sigma (σ,Σ) looks like O and E
@@JNCressey I'd also like to see what it would output if he tried to make the uppercase or lowercase versions of Greek letters, accented characters, etc. while the training data still only includes the Latin alphabet. :)
Saw the first two minutes of the video. Thought to myself: "Would make a perfect SIGBOVIK paper". 3:50 "You can read the paper, if you want". Downloading paper from site linked in description. Caption of Figure 7: "Although the output barely resembles letters, it does have a certain wispy Rorschach aesthetic, like a collection of delicate moths pinned to paperboard, that one could consider framing or publishing in the proceedings of SIGBOVIK 2021". I furiously nod in affirmation and admiration.
@@tom7 In a video packed so densely with plot twists, borderline useful information, and far-reaching humor references that make me scramble for google every few minutes, there's always some easter eggs that are going to be missed by some folks but noticed by others, and I can't help but sit back and wonder which are the ones that I missed :) That's half the fun after all!
You've kind of reverse engineered the OCR from the DS game, Brain Age. The tool assisted speedrun of Brain Age breaks their OCR system by just drawing arbitrary dots to trigger a specific letter recognition. i love this video lol
@@erifetim Everything has two edges. Like knives. Maybe the cure for cancer will also be used to GIVE cancer. It's not a problem with science or scientists. It's a problem with human nature.
5 minutes in: Wow, this is pretty interesting stuff. Funny guy, too. 24 minutes in: This is literally one of the most hilarious and interesting videos I've ever seen. I cannot express how much I love this channel. This is the epitome of quality.
It always made me mad when caps lock was on, and I instinctively hit shift to capitalize something, and I got a lowercase letter instead. You are the hero we need but don't deserve.
I wonder how well this model would have turned up had it been trained to Upper/Lower cases of other bicameral writting systems (Armenian, Cyrillic, Greek, some variations of Georgian, etc). Some languages that use latin script adopt special characters that themselves have capital forms, that could also be used for the training.
Not even a little disappointed. This channel is the computer science comedy I've been looking for my whole life :) thanks for doing this things. Someone had to!
It didn't get the prettiest results, but I still admire how closely you follow your actual train of thought with a spirit of experimentation. It's kinda like a Ben Levin video where you explore the topic more than demonstrate it, which is a cool thing!
This is brilliant. I laughed my ass off and learned some pretty cool stuff while i was in awe by both the simplicity and complexity of your presentation. Cudos
Nouns in German are called "Nomen", which is Latin for "name". Names are often capitalized in other languages as well, so Germany is *literally* calling things by their name. I always thought it was a good use of capital letters and it seems kinda nice towards 'things', which is an odd thing to say. On the other hand, they can turn nouns into adjectives, in which case they are not capitalized and is the opposite of what English does (ie. country names).
"signed distance fields", or the old "blow it up, blur it, shrink it a bit, sharpen it" trick you figure out using an ancient version of Paint Shop Pro, to get almost exactly the same effect.
Near the end of the video, and new to this channel love the content! I had an idea, not sure if you have already done something like this but... I think it would be awesome if you could include something like the area of the letter (number of pixels) and possibly make another ratio/relation using it (so the difference in the area of a to A could affect the area of upperer case A). Additionally using the area "ratio" you could have upper and lower limits of the area the next upper/lower case and you wouldn't end up with dust and blobs at either end so fast. Especially if you were to make it so the ratio had diminishing returns (as well as something like reducing it by some constant just so you don't have tiny/large letters very rapidly if there is a big different between upper and lower case). So if "A" took up 100 pixels, and "a" took up 75 (probably not even close to accurate, just chose them for easier calculations), the ratio is 4:3, or approximately 33% more/less per set, the lowerer case "a" should be around 56 pixels and the upperer case "A" should be around 133 pixels (assuming you dont use a constant to reduce the effect of the ratio). Using some value to reduce the ratio, you could fiddle around with it until you had more tangible upperer/lowerer and possibly upperest/lowest cases. And one last thing, if you were to make it so the pixels had to be contiguous, or have no more than x sections (probably 2, so letters like lower case "i" and "j" still work). you might end up with something closer actual letters. At this point just kinda ranting up to take it further lets say you you reduce the ratio by 50% and further reduce it by 2.5% for each depth (only up to 100%, so in this case a depth of 20) you would end up with: Upperercase being approx 100 (normal uppercase), 115, 133, 152, 172, ..., 450 (depth of 20) Lowercase being approx 75(normal lowercase), 65, 56, 49, 44, ..., 17 (depth 20) The numbers I chose probably werent the best ones to illustrate my point but you get the idea, and are far better at maths so if you had a go at it, you'd figure out something that works.
Okay, here’s the thing. Ack. This combined with your anagraphs video would have been EXACTLY WHAT Î NEEDED to see when I was writing my senior paper 3 years ago (only?) I still regularly think about that project, such as it was, and wonder whether I owe it to myself and the world to give it a more serious go now that I’m not barely scraping by in a haze of senioritis/being engaged to the love of my life. These two videos answered my question succinctly: nope, you’ve got it covered! Haha! So thank you!
I will mention here the project was ostensibly about ambigrams like SWIMS, words that can be spun around 180° and still read the same, or better yet, have a secret second message when read upside down. I wanted to create an alphabet from which one could construct any word as an ambigram, that is, a set of letters that, when viewed upside down, can pass for other letters. So if you want something to noodle on... This sort of machine learning process is exactly the thing I had far too little grasp of to really make the project work.
Thinking of how I'd approach the problem from a design perspective, I'd have thought a better solution would be to reduce the letters to simple lines+splines to generated the absolute shape/network of the new letter. Ie a T is just two lines meeting at a single vertex, an R is two lines and a curve etc. Then have a second program that converts these new basic shapes for the UUcase/LLcase into the typeface you want, so that T might have serifs or not and we convert a lower-lower-case letter into its specific desired typeface
I feel that it might help to lower the amount of fonts to just fonts that are fairly similar. Perhaps to train for upper cases a lower case you could use similar fonts that are sort of upgrades or simplified versions or other fonts.
[spoilers??]
the reveal of turning this into yet another system for making chess algorithms was absolutely amazing, 10/10
hey it's the rhythm heaven guy
hey it's the rhythm heaven guy
heres tree
oh hey it's the certified cute fraud!
toki! mi lukin e sitelen tawa sina. ona li pona!
Sometimes I write code that I think is clever or funny or imaginative. Then I see Tom over here making AI generated fonts battle each other in Chess.
I was confused by your profile picture for a moment
@@christianremboldt7028 C
"So the obvious thing to do with this... is use it to generate 26 different chess playing algorithms."
_my sides_
Nobody ever beats me when I play my Z opening.
@@ddpp3492 Zergling rush gambit?
you joke about toggling Caps Lock to write capital letters every time, but that's actually how my history teacher used to write capital letters, and she got pretty efficient at it
that poor caps lock key
Well, the Caps Lock key *is* a pretty historical thing.
I had a roommate in college that did the same thing. Didn't matter how much you told him about the shift key, he'd still do the caps lock thing every time.
I taught myself to do that on purpose 'cause pressing Caps Lock and seeing that silly little light turn on felt stress-relieving
@@Damaniel3 my dad still does that, its a thing for all those "hunt and peck" typers that never really learned how to type correctly or with more than 1 finger :D
now you need to go full XKCD and create capital numbers
what do you mean? capital letters exi-oh
numbers
uhhhh
There are already upper- and lowercase letters
Numbers are already uppercase. When they were making telegrams, Morse, teletypes, punch cards and so on, it's always uppercase letters and numbers first and lowercase later added on (this is why ASCII is roughly in the order of control codes, numbers, uppercase, lowercase, with punctuation sprinkled in between)
Typographically, uppercase has primary features going from the baseline to the ascender, and no features down to the descender, while lowercase letters have primary features only up to the midline, and only minor features like stems up to the ascender and down to the descender. Numbers are all primarily from the baseline to the ascent, and thus are uppercase.
Thank you for coming to my TED talk.
@@SimonBuchanNz I find myself surprisingly inclined to agree. Now I'm left thinking about different ways lowercase numbers might look.
I always thought of old-style numerals as "lowercase numbers".
Ok this whole video is hilarious but I cracked up the most at the "Mister I machine" line.
If you hadn't already written this comment, I would have 😂
"2.5 on the Bristol stool scale" is a very underrated joke.
This whole presentation is fantastic. I kept laughing harder and harder.
Ha! Forgot to look up what that meant at the time. 2.5 does seem about right, too
I was too lazy to look this up at first, but your comment finally made me do it, and it is indeed an amazing joke.
That one cracked me up too.
You ruined my search history. Who knows what adds I am going to start getting now.
@@Dziaji man doesn't use adblock
I work in software... i dream of one day being able to just make 5+ different user interfaces for some random project... frankly, this guy really is just on a whole other level.
I know, when I saw this video I just saw like 6 weeks of full time work 😂 he’s unreal
he is the definition of a "super user"
Next, have each chess-playing algorithm create it's own unique font. I'm sure you could figure out *some* way to do it...?
"I'm sure you could figure out some way to do it...?" Just train another neural network to figure it out x) !
I imagine you could make a program to first take the input letter and generate a chessboard based on it, then have each chess algorithm play against itself from that position for a set number of turns and use that new board state as the output character. Actually designing or training the algorithm to create a board that leads to interesting results for each chess algorithm would be the hard part.
Or create it's own chess font
And reverse engineer a printer to let it do the computing (like they did at bell labs in order to print chess books)...
@@Scouarn This is the correct solution.
There needs to be a word for "Useless but Entertaining Experimental Computer Stuff" because that is one of my greatest joys in life. It's also just obscure enough to be kind of hard to find using conventional youtube search methods, and now I've found a channel that does some very excellent UbEECS and great art to boot. Absolutely earned a subscription.
Turning it into a chess algorithm that just wants to make pretty shapes was pretty incredible.
Just goes to show how arbitrary and re-interpretable games can be. You don't _have_ to maximise efficiency, or use out-of-the-box tactics, or even aim to win; if the true goal of Chess this entire time was just to make fun shapes, we'd never know about any of that, which is a poetic window into an alternate objective.
"You will fail because you merely aim to win, while I aim to make beautiful shapes." - A chess agorithm that optimizes for the board looking like the letter 'C', moments before giving up both of its bishops.
would've been better if it had done a game tree traversal to find the one most likely to create a C, currently it's easy for it to make a terrible move where the opponent can immediately ruin the C
@@Metallicity lmao
@@alex15095 Now I'm imagining the perfect "chess" algorithm that will only ever goad its opponent into making a C out of all the pieces, and if its opponent ruins the C, the algorithm just keeps pushing the board until it can only ever look like a C.
I just found out you're a contributor at DFX. I've used and loved your plugins for a few years now. You're both incredibly, and intimidatingly talented! Thank you!
Cool! Thanks and thanks for using those! Yes, actually we have been working on destroy fx for the first time in many years... some new releases soon :)
this is how we all should do "projects". honestly, the results were a disappointment, and yet, you turned it into a lot of fun.
very free form jazz
I mean thats how most projects go tbh, hard to make things work, especially when you expect them to lol
its the art of presentation
The results were exactly what is expected from such an endeavor. I wouldn't even know how to manually tweak the results to make them more correct.
This is science !
"But frankly these results are fully bonkers" is a line that makes me laugh every single time I hear it.
Glad to know that I would somehow manage to loose against the letter t in chess
Imagine losing to the letter F...
@@theodoremurdock9984 F
I think this is my favourite comment ever
@@theodoremurdock9984 But it would give you a better respect for chess.
lose*
It would be really cool to try this with Chinese characters.
It would be easiest to do it with just Traditional and Simplified characters (for example, "horse" was simplified from 馬 to 马).
However, all of the characters have been evolving even before that, and many can be traced all the way back to the bronze age; a few thousand can even be traced back on 3,000 year old oracle bones. And those kind of resemble stone-age pictographs.
So maybe you could come up with a program that takes any doodle of an object, and puts it through thousands of years of evolution to instantly make a modern character. And then maybe see what it would look like however many years in the future.
Kana would be interesting too! It could be trained to “hiraganaize” and “katakanaize”, and then let loose. You could do wild stuff, like ask it for the katakana version of the letter D. The letter case data could also be used to create a lowercase を. Kanji/hanzi could also be added into the mix too, cause I’m very interested what a capital 議 would look like.
The English alphabet is kinda the best case (ha) scenario for this, it has simple shapes and they're all very distinct, spatially speaking, from each other. Chinese (and similar languages) have extremely complex shapes (even Simplified Chinese) that have no significant spatial relationships. It would require a stupidly gigantic network and a massive amount of training time.
your sense of humor is mesmerizing
Mr. I machine
Pushing the saying "because we can" to the next level! These videos are so entertaining and somehow reassuring me in my fear of wasting my life...why? I don't know.
this was hilarious, i cracked up as soon as you started talking about turning it into a chess algorithm
I absolutely lost it with the chess algorithm. It was funny before that but that’s when I couldn’t keep it together anymore.
hi stick
@@mmmmmmmmmmmmm hello
These videos lull me into a false sense that I am actually engaging in worthwhile thought, and it is moments like the one you described that snap me back to reality and make me realize that I am fully down the rabbit hole of nonsense.
One of my favourite youtube channels for impossible to explain reasons. It answers my burning questions, like "what letter is best at playing chess?"
19:27 this is why you use the softmax function on your categorical outputs. It doesn't just constrain the outputs to be between 0 and 1, it also makes them sum to 1 (and does it in a differentiable way, so you can backprop across it). With softmaxed probabilities, maximizing the 'f' score is actually giving you the thing that is "most like an f while also being least like everything else".
F
makes me think the model could perform better with some deep learning tinkering - since he says "the layer is initialized to a diagonal" it also seems that he's using a dense NN instead of a CNN, for instance
@Wave I think he said it initialized to the identity matrix. Which confused me as well.. should be using random weights. And yeah a model which uses convolutions and deconvolutions for an encoder and decoder would probably be superior if he is not doing that. Although with input and output sizes of 36x36 it's probably not *that bad* of an approach for just screwing around. The model will definitely take longer to train and probably be significantly less efficient, but if you throw enough parameters at it, it would probably compensate somewhat for the bad architecture.. and that's all machine learning really is right? Why spend time using my brain when I can just upgrade my gpu?
Uhh yeah right uhuhu
@@derpnerpwerp the reason he initialized it to the identity matrix is because it was added during training and randomizing the weights would make all prior training useless
I mean this entirely positive, this video is somehow defeating my insomnia within the first 5 minutes. Organized, rhythmic, intriguing. Absolutely fantastic, I hope I fall asleep before finishing it so I can rewatch it tomorrow night. :)
I'M GOING TO NEED AN UPPESTCASE VERSION OF THE LIKE BUTTON TO PROPERLY EXPRESS MY FEELINGS ABOUT THIS VIDEO.
i think i have found the most underrated channel on youtube
This video marks the beginning of the suckerpinch cinematic universe. Upestcase and lowestcase letters are now part of the chess video. Amazing.
"this is either generated image features or bus seat upholstery or a Gram stain of the same"
I just subbed even harder
when you said logical conclusion, I was sort of expecting upper and lower case numbers?
regardless, this is my favorite time of year and you can do no wrong. I admire everything you create and I can't wait for whatever comes next!
Thanks Josh! There are in fact uppercase and lowercase numbers and punctuation in the fonts, although they are not particularly captivating, alas.
@@tom7 How about greek letters and other scripts?
@@tom7 I would like to see uppercase and lowercase games of Chess.
Uppestcase Hieroglyphics
One of the greatest slow burn comedy videos I've ever seen
Best notification of my entire year
Your voice sounds so knowledgeable I don’t recognize any of it as humor, just as enlightenment.
babe wake up new Tom7 video
edit: Holy fuck this one went so off the rails. Thank you for being you Tom
It Could also be called Supercaps, meaning you could have a supercaps lock button, or a Megashift button for your keyboard. But since you have lowestcase, you need to invent 2 other keys, called undershift and Lowcaps lock. I don’t know if a keyboard inventor will invent this, but it would be awesome.
I think it should be an analogue dial that serves as a signal strength modulator for the ML network :D
WHERE HAVE YOU BEEN!? We’ve missed you!
Turns out I was in this comments section the whole time!
@@tom7 A bit late
@@tom7 Blink thrice if you need a new chess algorithm that designs a way to get you out of here.
@@tom7 This comments sectionis golden!
I think using mathematical double-struck letters (The ones used for sets) would be a very good standard for Uppestcase because they're easily recognizable and at the same time are well set apart.
𝕐ES THAT IS A GREAT IDEA! 𝕋HEY NEED TO BE A LITTLE BIGGER THAN THESE ONES THOUGH!
i love the use cases you give for each font
Every time you upload, I think to myself, "wait, who?" Then I watch, and it's great, and then go back through your old videos, and remember why I turned on notifications in the first place. Always a pleasure.
I love this entire video. The chess algorithm and that 3d model threw me off guard.
"Bus seat upholstery or a gram stain of the same" is such a good line.
i forgot how amazing this video is
I pretty much just refresh the page all day every day just waiting for Tom to upload
click the bell! makes it so much easier to not miss
@@Nasiulciaa it took 3 hours to notify me and that's 3 hours too much
@@user-sl6gn1ss8p 3 hours, what a scam! Thanks for your enthusiasm (:
Okay now, what is the most suckerpinch way to get notified if he uploaded a video?
@@Suppenfischeintopf chess algorithm
I like how the video simply doesn't stop. It just keeps going
I aspire to be the kind of man that could make people feel the way the end of every tom7 video has me feeling, just genuine happiness and amazement at what I just witnessed.
"it doesn't care about chess, it just cares about making Cs"
Dude, I joined a chess club in high school for the extracurricular requirement. I know exactly how that AI feels.
2:14 Why do easy things in 90 minutes when you can spend weeks automating something inefficient that is also needlessly complicated?
This is the stuff I live and breath for!
rule 34.♤: there is an XKCD comic for every single situation. no exceptions.
Gotta respect this man's dedication. Instead of just using convolutions he invents an entire new errorfuntion for his fonts.
The chess thing destroyed me, I actually started laughing. Well played
I love how you take an odd machine learning problem and then go through a bunch of odd but genius solutions you come up with. Please make more of these types of videos.
As always, this was glorious, and the ending was just...so good.
I'm glad to see you back making things, and I'm very excited to see what madness you do next!
Me too! Thanks! :)
dude, your channel is severely underappreciated
I feel strongly that someone has trained an AI on my taste in youtube videos and asked it to create the perfect video
This guy is insane. You put way too much effort into this, and I love it!
You should've included other languages upper and lowercase letters in the training data, there's no possible way that would contribute to more complexity or confusion
It surely won't get confused with these :)
- Nu (ν,Ν) looks like V and N
- Eta (η,Η) looks like N and H
- Gamma (γ,Γ) looks like Y and R
- Upsilon (υ,Υ) looks like U and Y
- Mu (μ,Μ) looks like U and M
- Sigma (σ,Σ) looks like O and E
@@JNCressey I'd also like to see what it would output if he tried to make the uppercase or lowercase versions of Greek letters, accented characters, etc. while the training data still only includes the Latin alphabet. :)
Saw the first two minutes of the video. Thought to myself: "Would make a perfect SIGBOVIK paper". 3:50 "You can read the paper, if you want". Downloading paper from site linked in description. Caption of Figure 7: "Although the output barely resembles letters, it does have a certain wispy Rorschach aesthetic, like a collection of delicate moths pinned to paperboard, that one could consider framing or publishing in the proceedings of SIGBOVIK 2021". I furiously nod in affirmation and admiration.
The inclusion of Sans in the very last frame is.... just some sublime perfection.
Thanks for noticing! Sometimes I wonder with the little Easter eggs..?
@@tom7 In a video packed so densely with plot twists, borderline useful information, and far-reaching humor references that make me scramble for google every few minutes, there's always some easter eggs that are going to be missed by some folks but noticed by others, and I can't help but sit back and wonder which are the ones that I missed :) That's half the fun after all!
This entire video I've been seeing the potential of using these fonts in games. So I'm going to do it. I actually LOVE perfect hallucination.
You've kind of reverse engineered the OCR from the DS game, Brain Age. The tool assisted speedrun of Brain Age breaks their OCR system by just drawing arbitrary dots to trigger a specific letter recognition. i love this video lol
I was thinking the exact same thing.
I think I've seen that speedrun. Was that in AGDQ?
I didn't use dots back when I used it, but short strokes.
After a while you could find these input shortcuts, and get very good times in Math.
I keep returning to this (and the chess) video at least once a year. This is just pure perfection.
Thanks! (:
This is exactly what I needed right now. Thank you.
17:19 "This font is based on Helvetica, which means 'of hell'" Holy crap that was awesome!
At some point, language models will be capable of generating research like this. Thrilling and scary.
When scientists did it because they could, but didn’t stop to consider if they should
@@erifetim Everything has two edges. Like knives. Maybe the cure for cancer will also be used to GIVE cancer. It's not a problem with science or scientists. It's a problem with human nature.
"either some generated image features or bus seat upholstery or a gram stain of the same" what a line! i subbed :)
3 and a half minutes in and I’m hooked.
I'm really diggin' the church Slavonic stylings of some of your letters. very cool. I too like to draw lettering, and also sea creatures
Your videos are the primary reason I look forward to April Fools every year (with the understanding there might not be one).
i have no idea how i stumbled upon this vid but omg i love it it just keeps taking one turn after another
oh my gosh "mr i"
TOM KNOWS HOW TO HAVE FUN!!!
5 minutes in: Wow, this is pretty interesting stuff. Funny guy, too.
24 minutes in: This is literally one of the most hilarious and interesting videos I've ever seen.
I cannot express how much I love this channel. This is the epitome of quality.
At about 21:00 I was literally screaming "WHY? why would you do it?" at the monitor. Amazing video!
And then you just do 3d models out of it.... for fun. As a person who dabbled in things like that, your programming skills are rally impressive.
It always made me mad when caps lock was on, and I instinctively hit shift to capitalize something, and I got a lowercase letter instead. You are the hero we need but don't deserve.
yessss i was just thinking about how much i wanted a new tom 7 video
Your channel is criminally small for how awesome and genuinely interesting the content is
Are you accusing me of a criminal endeavor
@@tom7 I plead the fifth
I wonder how well this model would have turned up had it been trained to Upper/Lower cases of other bicameral writting systems (Armenian, Cyrillic, Greek, some variations of Georgian, etc). Some languages that use latin script adopt special characters that themselves have capital forms, that could also be used for the training.
cyrrilic would be really uninteresting tho
Not even a little disappointed. This channel is the computer science comedy I've been looking for my whole life :) thanks for doing this things. Someone had to!
Mr I machine gave me a good chuckle :)
I love your sense of humor! The conclusion of this had some serious chaotic energy.
It didn't get the prettiest results, but I still admire how closely you follow your actual train of thought with a spirit of experimentation. It's kinda like a Ben Levin video where you explore the topic more than demonstrate it, which is a cool thing!
For some reason, the "Mister I" joke was the one that got me the most. Great video!
Yo I’ve literally thought about this but I never thought any mad lad would actually make it!
I had no idea what I was in for when I clicked on this video, but that was way better than anything I could have thought of.
the chess part turned this video from great to unbelievable
This channel is so underrated. I'm so happy to be coming across it now. I can't belive he doesn't have over 1 million subs already
Fun project and great video!
This is brilliant. I laughed my ass off and learned some pretty cool stuff while i was in awe by both the simplicity and complexity of your presentation. Cudos
this feels weirdly like purposefully failing to solve a captcha. It's like, "there are letters I can read, can you figure them out but also less"
I think this is the best UA-cam video I've ever seen
just rewatched your chess video the other day and was hoping for another april fools upload. thank you!
your playful attitude toward your projects is absolutely inspiring. thanks for helping me stay creative.
I must say, I really love the way you make things. Keep it up!
hi tux
@@elemenopi9239 bro go away
Every video you release is a gem. This is my favourite youtube channel by leaps and bounds.
I'm glad I'm not the only one whose made the "caʒ" joke before
This is the best thing the algorithm has ever recommended me for 10 days in a row
Nouns in German are called "Nomen", which is Latin for "name". Names are often capitalized in other languages as well, so Germany is *literally* calling things by their name. I always thought it was a good use of capital letters and it seems kinda nice towards 'things', which is an odd thing to say.
On the other hand, they can turn nouns into adjectives, in which case they are not capitalized and is the opposite of what English does (ie. country names).
Subscribed as soon as I finished the video, I hope there will be another before 2023 though, youtube seriously needs more creators like you
Tom7: 'We can also draw some weird shapes...'
Greek Eta: 'Am I a joke to you?"
"signed distance fields", or the old "blow it up, blur it, shrink it a bit, sharpen it" trick you figure out using an ancient version of Paint Shop Pro, to get almost exactly the same effect.
Near the end of the video, and new to this channel love the content! I had an idea, not sure if you have already done something like this but...
I think it would be awesome if you could include something like the area of the letter (number of pixels) and possibly make another ratio/relation using it (so the difference in the area of a to A could affect the area of upperer case A).
Additionally using the area "ratio" you could have upper and lower limits of the area the next upper/lower case and you wouldn't end up with dust and blobs at either end so fast. Especially if you were to make it so the ratio had diminishing returns (as well as something like reducing it by some constant just so you don't have tiny/large letters very rapidly if there is a big different between upper and lower case).
So if "A" took up 100 pixels, and "a" took up 75 (probably not even close to accurate, just chose them for easier calculations), the ratio is 4:3, or approximately 33% more/less per set, the lowerer case "a" should be around 56 pixels and the upperer case "A" should be around 133 pixels (assuming you dont use a constant to reduce the effect of the ratio). Using some value to reduce the ratio, you could fiddle around with it until you had more tangible upperer/lowerer and possibly upperest/lowest cases.
And one last thing, if you were to make it so the pixels had to be contiguous, or have no more than x sections (probably 2, so letters like lower case "i" and "j" still work). you might end up with something closer actual letters.
At this point just kinda ranting up to take it further lets say you you reduce the ratio by 50% and further reduce it by 2.5% for each depth (only up to 100%, so in this case a depth of 20) you would end up with:
Upperercase being approx 100 (normal uppercase), 115, 133, 152, 172, ..., 450 (depth of 20)
Lowercase being approx 75(normal lowercase), 65, 56, 49, 44, ..., 17 (depth 20)
The numbers I chose probably werent the best ones to illustrate my point but you get the idea, and are far better at maths so if you had a go at it, you'd figure out something that works.
Okay, here’s the thing. Ack. This combined with your anagraphs video would have been EXACTLY WHAT Î NEEDED to see when I was writing my senior paper 3 years ago (only?) I still regularly think about that project, such as it was, and wonder whether I owe it to myself and the world to give it a more serious go now that I’m not barely scraping by in a haze of senioritis/being engaged to the love of my life. These two videos answered my question succinctly: nope, you’ve got it covered! Haha! So thank you!
I will mention here the project was ostensibly about ambigrams like SWIMS, words that can be spun around 180° and still read the same, or better yet, have a secret second message when read upside down. I wanted to create an alphabet from which one could construct any word as an ambigram, that is, a set of letters that, when viewed upside down, can pass for other letters. So if you want something to noodle on... This sort of machine learning process is exactly the thing I had far too little grasp of to really make the project work.
Thinking of how I'd approach the problem from a design perspective, I'd have thought a better solution would be to reduce the letters to simple lines+splines to generated the absolute shape/network of the new letter. Ie a T is just two lines meeting at a single vertex, an R is two lines and a curve etc. Then have a second program that converts these new basic shapes for the UUcase/LLcase into the typeface you want, so that T might have serifs or not and we convert a lower-lower-case letter into its specific desired typeface
I never thought that by the end of this video I would be laughing out loud, but the high-quality render on the salt plains got me good.
I feel that it might help to lower the amount of fonts to just fonts that are fairly similar. Perhaps to train for upper cases a lower case you could use similar fonts that are sort of upgrades or simplified versions or other fonts.