"Uniqlo x Akamai sells another design of shirt in the same range which is plainly incomplete"
Imagine having to return a t-shirt because that malfunction!
— I don't understand why are you returning this, was the size wrong or you didn't like it?
— No, there is a syntax error at line 37 that makes it impossible to run, and I'm concerned people on the street may think I promote unsafe bash scripting.
I love this shirt! Here's a nice video from the actual designer about the process of making this shirt (including intentionally making it hard to OCR): https://youtu.be/jocGLiecpjU?t=526
Author here. Thank you so much for the link which I hadn't seen! I'm very happy to see this and I'm gratified that it was deliberately difficult to OCR, not just me.
OCRing this is a nightmare and is a good benchmark to any self-proclaimed good OCR/vision model.
I think though it could likely be easily OCR'd if you give the image to any decent agentic harness with a good vision model, e.g. newest Claude/GPT ones, and tell them to split the image per lines, and then just OCR each line individually.
I wonder if the script itself was written by an LLM before obfuscation? There seem to be a lot of comments in it, but in this case it's still ok :)
I don't think it was written by an LLM, some things stand out:
The congratulations text is both in English and Japanese. Contains a single heart emoji.
There was an intention to have a cyan to orange gradient, but the range starts in an ANSI block, ends halfway through the 256 color block and 256 terminal colors are not arranged like a gradient at all.
There's no sleep at the end of the loop where I feel like an LLM would add that defensively.
Safari's copy-text-from-image feature manages the entire base64 part of the string, except for the first character (I instead of a T). Weirdly, it gets much worse performance if you try to copy the entire string, including the hashbang part.
I wonder what it's doing under the hood to get such good performance?
I gave the photo to Opus 4.8 and it reconstructed the same script in one shot. Although it did say it had to correct some parts of it based on context where it suspected OCR mistakes.
> I wonder if the script itself was written by an LLM before obfuscation?
From the prototype shown here [0], and the way they talk about their process, I sincerely doubt it. Especially as they mention trying to make it hard for AI to handle the output.
I watched that whole video link - thank you for that - and he doesn't really say. In fact, he spends much more time on the beige color harkening to computer case plastics of the 80s & 90s.
What made me wonder, personally, was that the output seems identical if you use "♥PEACE♥FOR♥ALL" instead of the version with internal repeats. IF there is any point to that "manual expansion of the cycles", IMO that deserves a comment much more so than "# Calculate length of text; text_length=".
Also, that `echo -n ...` followed by `echo ""` instead of just plain `echo` in the first place seems like the kind of copy-pasta code LLMs generate. Then again, regular devs also write pretty bad copy-pasta code.
There is also this the weirdly "broken down" calculation with 3 `bc` invocations not 1 as if it was translated from a language with more arithmetic/special function power than bash.
There is also the color scale stuff done in the loop instead of outside (except the one color=$(..)) which seems very unnatural and also very like machine translation.
Also, at least for me, on my bash-5.3.15(1), `char="${text:t % text_length:1}"` does not work to slice out the multi-byte UTF8 heart symbols, but it sure does look like the kind of thing an LLM would do translating from a python3 script (such as something like https://news.ycombinator.com/item?id=48830669) into bash.
Another thing is, as others here have observed, there is nothing "gradual" about the xterm-256 color cube. So, "gradient" is a misnomer and exactly the kind of weird things LLMs do when they cobble text together.
Finally, all the tput stuff the script does instead of just "print x spaces" really smells like a human description of the side scroll in the video game graphic he shows inspired him somehow LLM-corrupted/complexified into the vertical scroll terminals do, but with a lot of extraneous complexity.
None of this is conclusive, but the video mentions 2023..2025 as when he did it and given that he was a designer and his concerns more visual than code-oriented, I'd have to say I disagree with your sincere doubt and I do strongly suspect the decoded script was very likely LLM-circa2024-generated, possibly with light post-edits.
The AI not handling the output relates to the final base64 output on the T-shirt (which other comments in this thread mention manually keying in or TFA discusses in the context of OCR). So, that is just not relevant to the question.
The last time Internet people were obsessed with OCRing some base64 was a few months ago when the DoJ released tons of emails from some guy who died, but they were released as rasterized PDFs.
Can't remember his name now, there's been so many distractions...
>I wonder if the script itself was written by an LLM before obfuscation?
I seem to recall seeing an Akamai-branded base64'd shell script on a white shirt pre-2021(?), so unless they've changed the code since then, I doubt it...
My code usually clocs at 50/50 (or thereabouts)[0]. Has, since my very first real engineering project (in 1987)[1]. I discuss in detail, here[2].
But one reason that I like LLMs, is that they help me to write even more documentation. I have found that I can instruct an LLM to revise my documentation, and make it even more effective.
[0] https://github.com/ChrisMarshallNY (My GH profile. Pretty much everything there, is like that -has, since long before LLMs were a broken rubber on the drug store shelf).
Human could write that many comments to get enough base64 text for a design. Maybe to even get some of the highlighted characters in places they want (roughly equally spaced apart).
Especially in a case like this, I would definitely write a lot of comments to aid in understanding, thus increasing trust so people would try it out and tinker with it.
Honestly it's a bit of a shame. I checked and they could've shortened their base64 payload by 304 chars by removing all comments except the top two congratulatory ones, or by 524 if they removed those too.
Would they still get the highlighted "PEACE FOR ALL" text throughout the shortened string? It looks like the length, and presence of those characters, was an explicit design choice.
Basically it just clusters same characters and asks the human to find the problems, which is easy when you're looking at a series of pictures like ssssss5sss.
The UI is kinda least-effort. Should ask a modern AI agent to make it look nice and intuitive, sometime maybe.
I thought it was funny that the author used a variety of OCR tools with mixed success before spending a lot of time manually fixing up the output from the best one, rather than just typing it in
That was also my thought… but I grew up mashing rubber keys for hours copying “games” out of magazines and books! Then hours after fixing all the typos!
Took me almost 2 minutes for 4 lines (and I missed a character in one of them!).
I would opt for OCR too, obviously so I'm prepared for the next bash t-shirt I'd come across...
I think this is a case where two people can successfully complete the task manually faster than one attempting to automate it. Get a ruler, read five centimetres of characters to your colleague, have them type it in as you go, then repeat that five centimetres back to you. Correct as you go. Format your string with the same line-breaks as the t-shirt, and remove them at the end, so you can be sure you've got the correct length on each row. Trial-and-error adjust the five-cm distance depending on your success rate as you go along
All in, you should have a non-corrupted string in 10-15 min.
I ran it through paddle paddle OCR and it flawlessly did it.
Google's OCR through my phone's Google lens had also worked at getting a very good extraction but not 100% correct. Definitely would spend less time fixing it than hand copying.
IDK what the author was using but I feel like he could have shared how his OCR attempt went, but I am thinking he tried some naive OCR tools.
Author here - that's a good idea actually, it shouldn't be too hard to compare the various attempts. The tools I used were whatever my Android built-in is (likely Google Gemini, but I can't tell whether this is something Samsung has replaced in OneUI); tesseract; tesseract with various tweaks and charsrt restrictions; Claude; and finally, manual fixes based on disagreements between all the previous.
Gemini3.5 Flash didn't have a problem OCR'ing and base64 decoding it, despite the OCR step having errors, it just fixed them in the base64 decoding step.
Great post! It's interesting, detailed but concise, and well-written. Also, I appreciate the "no cookies or tracking" and attractive, functional and performant site design.
> I guess Uniqlo is run through Windows though: one thing that struck me was the font, which I’m almost certain is Consolas,
Surely this would use whatever font the virtual terminal profile was set to? I don’t know of any method to choose a virtual terminal font from bash and don’t see any code that addresses it?
Author here. All hallucinations are my own. Now you point it out, I see why the jump in context from the terminal back to the tshirt font would give the wrong impression.
For anyone that cares, this is a slightly less stupid Python version:
#!/usr/bin/env python3
from os import environ; E = environ.get
from math import sin
from time import sleep
text = "♥PEACE♥FOR♥ALL" # The text to sine-scroll animate
nText = len(text) # Number of utf8 chars
freq = 0.2 # Frequency scaling factor
color0 = 12 # xt256 Color cube segment 12..<208
color1 = 208; nColor = color1 - color0
(w, h) = (int(E("COLUMNS", 80)), int(E("LINES", 24)))
t = 0
while True:
x = (w/2) + (w/4)*sin(t*freq) # x pos via sine value
x = max(0, min(w - 1, int(x + 0.5))) # bound to tty width
color = color0 + ((nColor*t)//h)%nColor # cycle colors
ch = text[t%nText] # Get char & Use xterm-256 color escs
print("%*s\033[38;5;%sm%s\033[m\n" % (x, "", color, ch))
t += 1
sleep(0.1) # original used bc shell outs to rate-limit
On one hand it's nice how it's clean and commented, but on the other hand some golfing could have made the encoded block a lot more reasonable to actually manually enter.
Might have to do something like that for a verse on the next Carolina Code Conference shirt. Been trying to figure out a good way to pull in cybersecurity.
Cool! I bought one a few months ago as soon as I spotted it at a Uniqlo store, and later ordered a larger size online—I really love wearing them. But it never occurred to me to look into the story behind them.
P
./cool.sh: line 31: bc: command not found
./cool.sh: line 34: bc: command not found
./cool.sh: line 37: bc: command not found
E
./cool.sh: line 31: bc: command not found
./cool.sh: line 34: bc: command not found
./cool.sh: line 37: bc: command not found
> I ran OCR in a few ways: First, using the built-in OCR of the circle-to-search feature on Android, which is often very good. Second, by using Tesseract with a few options and tweaks. And third by running it through Claude. After diffing the three to look for mismatches and getting Claude to output a table of locations for quick scanning, it became trivial but time-consuimg to tidy up the remainder
I bet 10$ I'd spend less time typing it from the t-shirt. And I wouldn't boil two kettles of water in the process.
But hey, AI makes you 10x more productive, I suppose
(Author here) for unrelated reasons my typing is very slow at the moment, so I was keen to automate. I see that people are getting different results from Claude than I did though.
Imagine having to return a t-shirt because that malfunction!
— I don't understand why are you returning this, was the size wrong or you didn't like it?
— No, there is a syntax error at line 37 that makes it impossible to run, and I'm concerned people on the street may think I promote unsafe bash scripting.
Upon seeing this, I decided to golf and came up with a shorter version:
I think though it could likely be easily OCR'd if you give the image to any decent agentic harness with a good vision model, e.g. newest Claude/GPT ones, and tell them to split the image per lines, and then just OCR each line individually.
I wonder if the script itself was written by an LLM before obfuscation? There seem to be a lot of comments in it, but in this case it's still ok :)
The congratulations text is both in English and Japanese. Contains a single heart emoji.
There was an intention to have a cyan to orange gradient, but the range starts in an ANSI block, ends halfway through the 256 color block and 256 terminal colors are not arranged like a gradient at all.
There's no sleep at the end of the loop where I feel like an LLM would add that defensively.
In my mind an AI would do something the most popular way even when that's not appropriate.
A human might do things in an unpopular way even when that's not appropriate.
I wonder what it's doing under the hood to get such good performance?
Looked it up, you put mouse over text, then just select and copy it - very cool!
https://support.apple.com/en-gb/guide/safari/ibrw20183ad7/ma...
On meetings I will often take a screenshot of the URL someone is presenting. I’m then able to just open the image and click the URL in the image.
Preview has pretty good background removal.
Notes will transcribe audio from audio files.
From the prototype shown here [0], and the way they talk about their process, I sincerely doubt it. Especially as they mention trying to make it hard for AI to handle the output.
[0] https://youtu.be/jocGLiecpjU?t=567
What made me wonder, personally, was that the output seems identical if you use "♥PEACE♥FOR♥ALL" instead of the version with internal repeats. IF there is any point to that "manual expansion of the cycles", IMO that deserves a comment much more so than "# Calculate length of text; text_length=".
Also, that `echo -n ...` followed by `echo ""` instead of just plain `echo` in the first place seems like the kind of copy-pasta code LLMs generate. Then again, regular devs also write pretty bad copy-pasta code.
There is also this the weirdly "broken down" calculation with 3 `bc` invocations not 1 as if it was translated from a language with more arithmetic/special function power than bash.
There is also the color scale stuff done in the loop instead of outside (except the one color=$(..)) which seems very unnatural and also very like machine translation.
Also, at least for me, on my bash-5.3.15(1), `char="${text:t % text_length:1}"` does not work to slice out the multi-byte UTF8 heart symbols, but it sure does look like the kind of thing an LLM would do translating from a python3 script (such as something like https://news.ycombinator.com/item?id=48830669) into bash.
Another thing is, as others here have observed, there is nothing "gradual" about the xterm-256 color cube. So, "gradient" is a misnomer and exactly the kind of weird things LLMs do when they cobble text together.
Finally, all the tput stuff the script does instead of just "print x spaces" really smells like a human description of the side scroll in the video game graphic he shows inspired him somehow LLM-corrupted/complexified into the vertical scroll terminals do, but with a lot of extraneous complexity.
None of this is conclusive, but the video mentions 2023..2025 as when he did it and given that he was a designer and his concerns more visual than code-oriented, I'd have to say I disagree with your sincere doubt and I do strongly suspect the decoded script was very likely LLM-circa2024-generated, possibly with light post-edits.
The AI not handling the output relates to the final base64 output on the T-shirt (which other comments in this thread mention manually keying in or TFA discusses in the context of OCR). So, that is just not relevant to the question.
Can't remember his name now, there's been so many distractions...
I seem to recall seeing an Akamai-branded base64'd shell script on a white shirt pre-2021(?), so unless they've changed the code since then, I doubt it...
My code usually clocs at 50/50 (or thereabouts)[0]. Has, since my very first real engineering project (in 1987)[1]. I discuss in detail, here[2].
But one reason that I like LLMs, is that they help me to write even more documentation. I have found that I can instruct an LLM to revise my documentation, and make it even more effective.
[0] https://github.com/ChrisMarshallNY (My GH profile. Pretty much everything there, is like that -has, since long before LLMs were a broken rubber on the drug store shelf).
[1] https://littlegreenviper.com/wp-content/uploads/2022/07/TF30... (Downloads a PDF)
[2] https://littlegreenviper.com/leaving-a-legacy/
Especially in a case like this, I would definitely write a lot of comments to aid in understanding, thus increasing trust so people would try it out and tinker with it.
The comments can be more cute/awe inspiring for people who aren't as familiar with bash but like solving puzzles as well.
[0] https://github.com/GL-Kageyama/UNIQLO_Akamai_T-shirt_Code
Basically it just clusters same characters and asks the human to find the problems, which is easy when you're looking at a series of pictures like ssssss5sss.
The UI is kinda least-effort. Should ask a modern AI agent to make it look nice and intuitive, sometime maybe.
All in, you should have a non-corrupted string in 10-15 min.
IDK what the author was using but I feel like he could have shared how his OCR attempt went, but I am thinking he tried some naive OCR tools.
> May the m×s/t² be with you
Surely this would use whatever font the virtual terminal profile was set to? I don’t know of any method to choose a virtual terminal font from bash and don’t see any code that addresses it?
Might have to do something like that for a verse on the next Carolina Code Conference shirt. Been trying to figure out a good way to pull in cybersecurity.
It’s a movie plot.
That's better than half the tech howtos out there.
The old one was: https://www.uniqlo.com/us/en/products/E459561-000/00
The guy who founded the new political party Team Mirai also wears it frequently: https://en.wikipedia.org/wiki/Takahiro_Anno
https://www.wired.com/2000/08/court-to-address-decss-t-shirt...
Very wow. Shame they assumed everyone has "bc"...
I bet 10$ I'd spend less time typing it from the t-shirt. And I wouldn't boil two kettles of water in the process.
But hey, AI makes you 10x more productive, I suppose
My bet is against manual OCR with various engines + finding mistakes later using an LLM.