Jonathan Hoefler on type ai (on threads)

Dave Crossland

https://www.threads.net/@jonathanhoefler/post/CzvC74wOre6

Thoughtful thread by Jonathan Hoefler 

Typofactory

interesting thoughts. My first inclination is that the reason LLM and image-generating AIs being so “good” says more about the tasks themselves than the skill of AI. Once it becomes “easy” to generate fonts, I would expect an influx in AI (and low-quality) generated fonts. Type design is a very detail-oriented craft, and, (smooth transition to my own worldview) I think that AI will be similar to the calculator: a large change, but not paradigm shifting.

Enrico Sogari

@Dave Crossland Thanks for the heads up.

Vasily Draigo

I don't think AI should replace the design process of creating a font. A type designer can always make a prototype, it's a different process for everyone, it's the creative part of creating a font. Then you can give this prototype to the AI to create variations and the variation you like is just finalized. We have a regular weight, here and the main force of AI will enter the technical part, thanks to AI will be able to make high-quality interpolation with only one weight, of course with settings and variations. And further, spacing, any open type features, and even the great and terrible hinting)). We still need to make one quality font, but that's what font designers are for.

Jamie Clarke

The range of opinions for where AI takes humanity is enormous: from utopia to annihilation. However, for type design it seems the options are coalescing around two outcomes:
1. Totally AI generated fonts (fonts-on-the-fly): Free and ubiquitous, likely of dubious quality (to begin with) but won't deter most consumers. Not a good outcome for type designers.
2. AI as a co-pilot to do the heavy lifting in type design - 'make the bold', 'give me options for a true italic', sort of thing. Great outcome for type designers.

Are there other outcomes that have a comparable level of consensus?
Will AI inevitably revert to option 1, even in a utopian outcome?

Ray Larabie

3. Fonts are no longer needed.

I foresee a fascinating shift in how we approach typography, particularly with the rise of machine-generated lettering as opposed to fonts or typefaces. Imagine a world where lettering isn't just static, but dynamically adapts to context, providing a level of flexibility and precision that goes beyond what current fonts offer*. There would be no need for fonts to be generated, internally or otherwise.

This advanced machine lettering would be capable of understanding and adjusting to the nuances of a message, much like how ChatGPT can adapt its tone from formal to casual, or even sarcastic. It would cater not just to the content, but also to the preferences and needs of the reader. For instance, someone with vision challenges might see a different, more accessible version of the text compared to someone with clear eyesight.

The beauty of this technology lies in its ability to make every line of text unique and contextually relevant, akin to the work of a skilled sign painter or calligrapher. This means perfectly justified lines and a harmonious blend of style and substance, where the lettering complements both the message and the surrounding visual elements.

As this technology matures, the use of conventional fonts might gradually become more of a historical fascination than a necessity. The quaint charm of fixed fonts may still hold appeal, but their role could shift to being more about preserving a rich typographic heritage rather than serving practical needs.

It's an exciting time for typography, with boundless possibilities on the horizon. I’m not super thrilled about how it will affect my business, but I think it’s pretty neato.

* You don’t have to try too hard to imagine that because I think I’m describing what scribes did.

Jamie Clarke

Thanks Ray, that's really insightful. I can imagine it. I think this goes beyond 'fonts-on-the-fly' towards ever more specialised contexts.
I also agree with your last sentence!

Enrico Sogari

I'm new to type design, I'd been working as a web and brand designer for a couple of decades and then moved to type design just a few years ago. So even though I'm not a young designer I'm new to this and I have to say, although the future of type design looks very precarious right now because of AI, I don't think I can stop doing this. I enjoy it too much, it completes me like anything else before, professionally speaking. I think AI (starting from 2024) will be brutal for this space but there must be a way to safeguard the craft.

Jamie Clarke

Thanks for saying that Enrico. I feel much the same about type design and I made the switch from Graphic/Web design a decade ago. 

There's no stopping the technology, but I think the dilemma in all creative industries is that AI can be used to enhance our design work and take us to new heights, or encourage a flood of mediocre design.

Now more than ever, good taste, judgement, and craft must steer the use of the technology. AI is already democratising design, so professionals must set the bar high as an example for enthusiasts. I feel it's our duty as designers to ensure that AI is steered not just toward improving efficiency but more importantly the quality of our work.

Dave Crossland

Ray, I think calling automated lettering is a great idea. I see the same end game: fonts are obsolete, there's just documents, and the contents of the document, text of all kinds, photos, illustrations, iconography, typography, layout, and typefaces (well, now it's lettering) all get generated in milliseconds, and can be regenerated and refined over and over until the author is satisfied. The amount of text you bring with you could be a full draft or a sentence prompt, same as the amount of photos, layouts or type.

Craig Eliason

Particularly with spacing, fonts are by necessity full of compromises (even with kerning). Because fonts presumably dominate the lettering in training data for AI, I imagine there’s this funny effect where AI lettering isn’t theoretically fettered to those compromises when generating text, but it doesn’t know better!

Ray Larabie

@Dave Crossland I think you're right on target, but the one thing that's missing in your scenario is consistency. However, I think this will be solved without fonts. One possibility is the use of short strings of characters, like simple hashcodes or prompts, to capture the essence of a particular lettering style. This approach would streamline the design process significantly. Imagine a scenario where a designer inputs a prompt into their design tool, generating several typographic options. Once a style is selected, the tool creates a hashcode for that specific typeface design. This code can be shared easily among designers or saved for future use, allowing for the recreation of the exact lettering style with minimal effort.

Another intriguing prospect, given the advancements in AI in November, involves the use of specialized AI instances. These AIs would be trained to master one specific typeface, understanding every nuance of its design. When a layout requires typographic elements, the primary design AI could connect to this specialist AI, functioning like a plugin. This setup ensures that any type used is consistently on-brand and visually correct, guided by the expertise of the specialist AI.

The particular advancement I'm referring to is training individual GPTs in ChatGPT, which was introduced this month. I believe there's a potential for linking these trained specialists together to consistently perform more complex tasks. You can think of the old model as being one smart, talented individual that can do everything and the new model as an office full of smart, talented specialists. When they linked together, they will act more like a company rather than an individual. Therefore, it would be simpler to have a full-time specialist adept at writing in a particular style, than storing a set of fonts.

Edit: I apologize for the length of this message, but I want to highlight that users can now make individual GPTs, including those skilled in visual tasks, publicly accessible. For instance, ChatGPT Plus members can access a GPT specialized in transforming faces into Simpsons characters. Envision a future tool capable of producing lettering, potentially leading to a new font market. Here, instead of crafting fonts in an editor, typeface designers would essentially train AI scribes. Modern type designers and typographers would produce the most talented of these scribes, with companies valuing quality and consistency in typography willing to pay for access to them.

Dave Crossland

Hashcodes could work, yes, but I think they will just be names: you just give the name of the typeface in the overall prompt, and that will be fine; or if the typeface has no name, just give the location of a prior document that has it, to use as a seed and ask it not to deviate much. 

The chat gpt stuff just seems like a gui for making a long initial prompt that's properly structured, given the very long prompt length now possible in the general model?

Ray Larabie

@Dave Crossland Seed number, not hashcode…that's what I was trying to think of. Some AI tools already use seeds to recreate image styles, but yes, there's no reason it couldn't be a typeface name.

You're correct, the instances which OpenAI calls “GPTs” are similar to using a long initial prompt. The main difference is that GPTs can have their own linked documents. For example, you can load text, PDFs, spreadsheets, photos, or other documents and instruct the GPT to use them as reference. Unlike the long initial prompt technique, you don't need to keep reminding it to stay on track. You can also modify the instructions later.

John Hudson

I wrote some speculative thoughts on machine typography as a coda to a chapter on typographic technology (not yet published). The thought I had is that the machine is likely to be very good at producing the kind of impersonal, conventional typography that I likened to the kind of that Tschichold was able to specify for typesetters to follow at Penguin. The conventions are easily modelled, amply demonstrated, and vary mainly according to media format. What I think the machine will not be good at—or rather, a thing that it will do only superficially, uncomprehendingly—is the kind of personal, individual typographic expression that typified Tschichold’s early, modernist work, where the result seems, inevitably, to be a kind of mechanical pastiche (as is the case with so much that machine learning is already producing based on prior human art).

Dave Crossland

I think impallari already showed us all here on this forum that is equally capable of producing minion as kablammo

John Hudson

I was talking about typography, not typefaces.

What Pablo and others have shown is that machine learning can produce pastiches of a wide range of things that have already been created by humans. Some things are more obviously pastiches, because what they mimic are idiosyncratic or personal, strongly identified with an individual human voice, while others are less obviously so because they mimic culturally broad and conventional works. I think Tschichold’s two modes of work do provide a useful way of thinking about this difference, not least because he identified the difficulty of the earlier mode: it wasn’t easy to do well and required a specific intelligence or feeling, while the later ‘classical’ mode could be accomplished by anyone with basic typesetting skills following a set of rules or models. Since the latter mode embodies those models, mimickry of that style of typography is indistinguishable from following the models, so machine learning looks like actual learning (understanding of concepts and principles, rather than mere repetition), while mimickry of Tschichold’s modernist work will look like stylistic pastiche, without any evidence of understanding. Arguably, that is in a sense more ‘honest’, since it doesn’t give the impression of intelligence.

Dave Crossland

I'm skeptical it will not be good at it.

https://twitter.com/hey_madni/status/1730561373346971995

John Butler

I take comfort in the fact that these AI design technologies seem most focused on generating the same kind of soulless throwaway visual pollution that for-profit art colleges used to train humans to churn out.

John Hudson

I'm skeptical it will not be good at it.

Let me clarify, or rephrase: it isn’t that the machine won’t be good at creating pastiches, but that creating pastiches is not a good thing. As I characterised it in the article I was writing: this is the bad thing that the machine is very good at.

Dave Crossland

Isn't this like saying type is a bad imitation of calligraphy, and our newspapers should be photocopied scribal pages?

John Hudson

Isn't this like saying type is a bad imitation of calligraphy, and our newspapers should be photocopied scribal pages?

No, because type is not a bad imitation of calligraphy. Even when it was most directly referencing scribal page layout and script structure, type was never simply pastiche but was always also exploring new directions, because of the conscious intelligence of the people doing the work, who were able not only to learn from what they had seen but to make conceptual advances. Machine learning is massively and rapidly productive, but is so far limited in terms of understanding of what it is learning and not even trying to make conceptual advances because current AI is not conceptual at all. It is still Eliza, just a lot more powerful and with a lot more content from which to learn what probablistically to say next.

Imagine giving an AI prompts that reference dada or surrealism. It will easily return pastiches of Duchamp or Dali—predictably, the better known artists associated with those movements—, but it won’t understand anything about dada or surrealism, so won’t have any conceptual means to apply the ideas of dada or surrealism in new ways. And I chose dada and surrealism as examples because, like Tschichold’s early modernist work, they are concerned with the unexpected—hence unpredictable—and with challenging conventional associations—hence non-probablistic.

Dave Crossland

John Hudson said:

Isn't this like saying type is a bad imitation of calligraphy, and our newspapers should be photocopied scribal pages?

No, because type is not a bad imitation of calligraphy. 

Tell that to the Urdu script readers in Pakistan

Personally, and with no special insight, I think you are going to be surprised how good the current generation of Eliza systems will get at producing what looks like abstract conceptual reasoning and doing things we would say are new if a person did them

Nick Shinn

My opposition to AI has nothing to do with pastiche vs. originality, or copyright.
It’s a step too far, as if the previous generation of digital technology hasn’t already proven to be a disaster for humanity.
It’s not about the quality of technology, but the asymptotic the rate of change, way too fast for us to mitigate, via legislation and adaptation, the disruption it causes.
And yet still we think more of the same will fix the problems it causes. Duh.

As a type designer, I am continuing to explore a feature of “old” tech — the contextuality of OpenType, and will have nothing to do with AI. Although I will no doubt be forced to use it, one way or another, by the planned obsolescence of automatic software upgrades, or become excluded from society.