While we're at it: my own work 2 years ago in creating an entire workflow for turning Midjourney or DALL-E dropcaps into attractive, lightweight, easy-to-create dropcaps for web pages: https://gwern.net/dropcap We use it for the cat, Gene Wolfe, and holiday pages.
This is pretty cool. Can't believe how many got I 100 just on my mobile. I consider myself horrible at fonts and design and could never design my own, but maybe good styling is something that can be internalised
Definitely not the first AI generated font. One can find an enormous amount of research in AI font generation on https://scholar.google.com/ going back many years. This could possibly be the first one that used Nano Banana though, and the result is impressive for sure!
I would also like to see a custom background image. And floating text or bubbles that follows your cursor. And maybe play your favorite song in the background. And a "best friends" list (oh substack already has that basically!)
> Copyright law does not protect typeface or mere variations of typographical ornamentation or lettering. A typeface is a set of letters, numbers, or other characters with repeating design elements that is intended to be used in composing text or other combinations of characters, including calligraphy. Generally, typeface, fonts, and lettering are building blocks of expression that are used to create works of authorship. The Office cannot register a claim to copyright in typeface or mere variations of typographic ornamentation or lettering, regardless of whether the typeface is commonly used or unique.
Given the incredible amount of work that goes into designing a typeface I find that really surprising.
Apparently you CAN protect the implementation of a typeface, e.g. the font file itself. Wikipedia says:
> Typefaces and their letter forms are considered utilitarian objects whose public utility outweighs any private interest in protecting their creative elements under US law, but the computer program that is used to display a typeface, a font file[a] of computer instructions in a domain-specific programming language may be protectable by copyright. In 1992, the US Copyright Office determined that digital outline fonts had elements that could be protected as software[13] if the source code of the font file "contains a sufficient amount of original authorship".
Likewise I think it is extremely dubious that models can be copyrighted at all, for the exact same reason you can't copyright a phonebook or database. The entire regime of claiming to release models under various licenses is bullshit, because you can't copyright rote transformations of things either.
It's completely different from a phonebook or database, which are mere compilations.
If something is considered sufficiently transformative, then it can be copyrighted. If you do a bunch of non-trivial processing on a database to generate something new, you can copyright that.
And LLM training is in no universe a "rote transformation". It is incredibly sophisticated, carefully tuned, and results in a final product that could not possibly be more different.
This is definitely going to be argued over in court at some point, along with many other questions about AI and copyright.
Speaking of which, why is it taking so long to get a supreme court decision on whether or not training counts as copyright infringement? The only court cases that have been resolved so far have settled on unrelated grounds without touching the core issue.
Not "world's first" by a long shot.
Someone's made a python script in June 2024 to do it semi-automatically using SD 1.5: https://github.com/414design/4lph4bet_processor
I'd like to add that tom7 used AI to generate an upperercase and lowerercase font in 2021. https://tom7.org/lowercase/
While we're at it: my own work 2 years ago in creating an entire workflow for turning Midjourney or DALL-E dropcaps into attractive, lightweight, easy-to-create dropcaps for web pages: https://gwern.net/dropcap We use it for the cat, Gene Wolfe, and holiday pages.
AI should play a few runs of this game before attempting font design:
https://type.method.ac/
This is pretty cool. Can't believe how many got I 100 just on my mobile. I consider myself horrible at fonts and design and could never design my own, but maybe good styling is something that can be internalised
This is fun. Thanks for sharing
I enjoyed this far too much!
> Worlds first AI generated font
hmm, not sure: https://www.linkedin.com/feed/update/activity:72925311263198... (disclaimer: i'm a former employee)
Definitely not the first AI generated font. One can find an enormous amount of research in AI font generation on https://scholar.google.com/ going back many years. This could possibly be the first one that used Nano Banana though, and the result is impressive for sure!
As opposed to putting an LLM inside a font
https://fuglede.github.io/llama.ttf/
Ok but what is the Russian Orthodox font at the top of the article?
why dont they show you it in use? seems like an obvious thing.
> Worlds first Ai generated font
For a brief moment I thought the title was referring to Adobe Illustrator.
The writing style of this was so chaotic. I loved it, in a loopy, end-of-day sort of way
> I also found out that my friend's company got charged $2,000 per character. WTF.
Is there more context to this? I can't see anything preceding it that explains what it's referencing.
Just the fact that a lot of people pay a lot of money to make fonts.
Design and branding agencies cost a lot
They are good for a cohesive message
But their utility has just plummeted
I'm pretty sure Grant Sanderson capitalizes his first name.
I’m sure Microsoft would love to include it in WordArt.
> Imagine each Substack owner can make their own font to highlight the essence of their writing.
No please, for the love of all that is good and holy, please no
I would also like to see a custom background image. And floating text or bubbles that follows your cursor. And maybe play your favorite song in the background. And a "best friends" list (oh substack already has that basically!)
And blinking too
A font without copyright is not a real font.
Typefaces cannot be copyrighted in the US, so that’s really irrelevant.
Huh, TIL: https://www.copyright.gov/circs/circ33.pdf
> Copyright law does not protect typeface or mere variations of typographical ornamentation or lettering. A typeface is a set of letters, numbers, or other characters with repeating design elements that is intended to be used in composing text or other combinations of characters, including calligraphy. Generally, typeface, fonts, and lettering are building blocks of expression that are used to create works of authorship. The Office cannot register a claim to copyright in typeface or mere variations of typographic ornamentation or lettering, regardless of whether the typeface is commonly used or unique.
Given the incredible amount of work that goes into designing a typeface I find that really surprising.
Wikipedia has some good coverage on this: https://en.wikipedia.org/wiki/Intellectual_property_protecti...
Apparently you CAN protect the implementation of a typeface, e.g. the font file itself. Wikipedia says:
> Typefaces and their letter forms are considered utilitarian objects whose public utility outweighs any private interest in protecting their creative elements under US law, but the computer program that is used to display a typeface, a font file[a] of computer instructions in a domain-specific programming language may be protectable by copyright. In 1992, the US Copyright Office determined that digital outline fonts had elements that could be protected as software[13] if the source code of the font file "contains a sufficient amount of original authorship".
Likewise I think it is extremely dubious that models can be copyrighted at all, for the exact same reason you can't copyright a phonebook or database. The entire regime of claiming to release models under various licenses is bullshit, because you can't copyright rote transformations of things either.
It's completely different from a phonebook or database, which are mere compilations.
If something is considered sufficiently transformative, then it can be copyrighted. If you do a bunch of non-trivial processing on a database to generate something new, you can copyright that.
And LLM training is in no universe a "rote transformation". It is incredibly sophisticated, carefully tuned, and results in a final product that could not possibly be more different.
This is definitely going to be argued over in court at some point, along with many other questions about AI and copyright.
Speaking of which, why is it taking so long to get a supreme court decision on whether or not training counts as copyright infringement? The only court cases that have been resolved so far have settled on unrelated grounds without touching the core issue.
This is a very strange perspective to take.
That's the joke
Pretty interesting that this much is possible. Too bad the fonts are terrible.
I had tried vector based work a couple years ago and generative AI models were very limited at keeping up
I could see them being better now, havent revisited
[dead]