Claude 5.5 creating fonts from handwriting?

Hello everyone,

Saw this recent tweet from Anthropic claiming they generated a fully working font from 120 pages of handwritten notes using Claude Opus 5.5. https://x.com/claudeai/status/2111177838208377213

Comments

  • Qwerasd
    Qwerasd Posts: 23
    It's not very good. Existing "handwriting to font" converters do a much better (though still generally flawed) job. A "working font" is an incredibly low bar, the actual quality of what was produced is terrible, I don't know how someone could be satisfied with that result... maybe the person who did it wasn't actually satisfied with the result but didn't want to keep spending more money on tokens for the slop machine.
  • "slop machine" is the perfect term for it. It really highlights the gap between simply mapping raw shapes to Unicode values and actual type design. It’s going to take a lot more than just processing power to teach AI about optical corrections and balance.
  • Dave Crossland
    Dave Crossland Posts: 1,586
    Yet another slop-font app came across my feed today: https://handset-867c3.web.app
  • "Claudicant" font :)
  • Mark Simonson
    Mark Simonson Posts: 1,812
    edited October 21
    Interesting that it failed to use the extremely wide word spacing of the handwriting. (Or, maybe this difference was specified in the prompt?)
  • pavljenko
    pavljenko Posts: 42

    If you look at exactly what they did, it doesn’t look so terrible. They essentially introduced a new way of rendering fonts — through code. And I believe this is the best thing that has happened to font design since the invention of digital fonts. In the near future, we’ll see a lot of new developments and reworks of beautiful fonts with perfect outlines and encoding. Handwritten font is just the beginning.

  • pavljenko
    pavljenko Posts: 42
    fonts have been rendered with code for about 60 years!

    Ray, you're right, I put it badly: I didn't mean rendering. I meant that the agent draws the letters in code. It wrote a program that finds the path of the pen in each letter and builds the outline around it, and nobody drew a single curve in an editor. Fonts made by code aren't new, of course; what's new is that the agent writes that code itself, from a task in plain words.

    One could try giving the agent exactly the task you describe: reconstruct the hand and the pen, then simulate the movement of the hand, the action of the pen and the friction of the paper, to get a more faithful font. But the task in this post was apparently something else.

    Translated by AI.

  • John Hudson
    John Hudson Posts: 3,879
    But ‘find the path of the pen in each letter’ is exactly what it did not do. It found the centre line of shapes in an image, and as Ray notes there is nothing new about this. It did not find the path, because it knows nothing about ductus. A lot of the paths it draws along the centre line happen to correspond to the ductus of the pen, but that is coincidence, not understanding. Many other paths it draws are not the way the shapes were written, and the AI cannot tell the difference, because all it has to work from are the shapes, not an understanding of how they were written. This gets towards the problem of instrumentalism that came up in the autokerner thread: the limitation of judging AI by its results rather than how it got to those results. With more processing power and better training, of course, the results keep getting better. Until they don’t.