It is argued that readers are likely to carry away from Donald Knuth’s article “The Concept of a Meta-Font” [Knuth1982] a falsely optimistic view of the extent to which the design of typefaces and letterforms can be mechanized through an approach depending on describing letterforms by specifying the setting of a large number of parameters.

Through a comparison to mathematical logic, it is argued that no such set of parameters can capture the essence of any semantic category.

Some different way of thinking about the problem of the “spirit” residing behind any letterforms are suggested, connecting to current research issues in the field of artificial intelligence.

…Interpolating between an Arbitrary Pair of Typefaces: The worst is yet to come, however. Presumably Knuth did not wish us to take his rhetorical question in such a limited way as to imply that the numbers 6 1⁄7 and 1⁄4 were important. Pretty obviously, they were just examples of arbitrary parameter settings. Presumably, if Metafont could easily give you a 6 1⁄7-point font that is 1⁄4 of the way between Baskerville and Helvetica, it could as easily give you an 11 2⁄3-point font that is 5⁄17 of the way between Baskerville and Helvetica—and so on. And why need it be restricted to Baskerville and Helvetica? Surely those numbers weren’t the only “soft” parts of the rhetorical question! Common sense tells us that Helvetica and Baskerville were also merely arbitrary choices of typeface. Thus the hidden implication is that, as easily as one can twiddle a dial to change point size, so one can twiddle another dial (or set of dials) and arrive at any desired typeface, be it Helvetica, Baskerville, or whatever. Knuth might just as easily have put it this way: “The ability to manipulate lots of parameters may be interesting and fun, but does anybody really need an X-point font that is Y% of the way between typeface T1 and typeface T2?”

…Thus we realize that Knuth’s sentence casually implies the existence of a “universal ‘A’-machine”—a single Metafont program with a finite set of parameters, such that any combination of settings of them will yield a valid “A”, and conversely, such that any valid “A” will be yielded by some combination of settings of them. Now how can you possibly incorporate all of the previously shown typefaces into one universal schema?

[See: font GANs interpolating in their latent space! As for Hofstadter’s ‘shattering’ argument, the fact that you can embed new images into GAN latent spaces usually without any problem would seem to disprove his claims there…]

…A letterform-designing computer program based on the above-sketched notions of typographical roles and niches would look very different from one that tried to be a full “mathematization of categories.” It would involve an integration of perception with generation, and moreover an ability to generalize from a few letterforms (possibly as few as one) to an entire typeface in the style of the first few. It would not do so infallibly; but of course it is not reasonable to expect “infallible” performance, since stylistic consistency is not an objectively specifiable quality. In other words, a computer program to design typefaces (or anything else with an esthetic or subjective dimension) is not an impossibility; but one should realize that, no less than a human, any such program will necessarily have a “personal” taste—and it will almost certainly not be the same as its designers’ taste. In fact, to the contrary, the program’s taste will quite likely be full of unanticipated surprises to its programmers (as well as to everyone else), since that taste will emerge as an implicit and remote consequence of the interaction of a myriad features and factors in the architecture of the program. Taste itself is not directly programmable. Thus, although any esthetically programmed computer will be “merely doing what it was programmed to do”, its behavior will nonetheless often appear idiosyncratic and even inscrutable to its programmers, reflecting the fact—well known to programmers—that often one has no clear idea (and sometimes no idea at all) just what it is that one has programmed the machine to do!

See Also: