Google Pics

I am always sympathetic to craft, and the point is certainly well-made that the code generated by current models of AI is based on training on vast amounts of code written by humans (and yet still manages to be ‘hideous and harebrained’ — perhaps these are the ‘transformative’ qualities that AI brings to its derivations from other people’s work and ideas?).

I think there remains a distinction, though, that is particular to software, which is that the value in software is in what it does, while the craft that goes into writing code is invisible to the vast majority of people who use the software. People are unlikely to see the beauty that you see, because they are unlikely to ever see the code.

Anything that human beings do can embody craft, and anything that human beings make can possess that extra-utile quality of beauty that we call art. We are a crafty, art-making species. But craft and art are often hidden and hence not valued in the utile valuation of things used.

I know some very good electricians, who practice their profession with skill and considerable knowledge and practical craft. If one were to strip away the drywall, one might even admire a kind of beauty in the way they have wired a house efficiently and tidily. But the inhabitants of the house only care that the light goes on when they flip a switch, and that their appliances work when plugged in. Not only are these people not aware of how well or poorly the house has been wired — presuming everything works —, they are mostly ignorant of electricity and would struggle to explain it to someone.

I have invented some clever ways to do things in OpenType Layout — programming of sorts, and something I’ve been told is ‘Turing complete’ —: someone looking at the source for my GSUB and GPOS tables might admire how I have handled kerning to zero-width marks in Telugu script, or contextual variation of ornamentation on otherwise identical words in some stylistic sets in the Gabriola display font. But most users of these fonts are never going to see those sources and, insofar as they use the fonts, only care what outcomes they produce. What they are going to see is the typeface design, which is the primary distinctive aspect of any font: the first reason they have chosen to use this font rather than another, and whose visible, graphical aspect is, hence, where the primary value of the font subsists. Other aspects of font functionality have value, of course: the character set coverage, the accessibility to all the forms needed to typeset a particular document in a particular language, etc., but all of these are experienced in the visible outcomes of using the font, not in how it was made or how the tables are structured and active programmed aspects were written.

This is true of all software I can think of, from fonts, to games, to any kind of app, to the back-end of a website: the value is in the use, not in the code. [This is why intellectual property law around software is messed up in the USA and many other places: it seeks to apply protection somewhere other than where the value subsists.]

Circling back to AI: I acknowledged above that the code generated by current models of AI is based on — i.e. looted from — vast amounts of code written by humans. That is how these models work, and it is very much in keeping with the kind of digital colonialism and extraction that characterises the whole baleful AI plunder of human creativity by capital. It seems to me, though, by way of a thought experiment, that a programming language, as a way of directly or indirectly instructing a computer to perform certain functions and tasks, could be taught to a machine from first principles of programming and then language-specific ways to do things, in the same way that a human being would learn to program. That’s not the quick-and-dirty (oh, so dirty!) kind of LLM approach in current use, but it doesn’t seem inconceivable in a different model. This is not to say that programming languages are fully determinative, such that there is only one way to affect a function or task in the computer in any given language, but they are more so than e.g. natural language in human communications, precisely because the latter is not directly functional: outside of the military, what someone says does not compel what another person does in the way that a command in a line of code compels what the computer does. So, yes, code can be beautiful and elegant, and code can be hideous and harebrained — and I have witnessed plenty of cases of human programmers criticising each others’ code, over the years —, but ultimately its value is going to subsist in its outcomes, not in how it was written.

That seems to me to be what makes it easier for people to think AI ‘is actually good if I use it for this other thing that I know little about but I find it useful for’, when that thing is writing code. And I can totally see why that is frustrating to people who have found humanly fulfilling experience in learning to program and in the practice of doing it well, for whom that fulfillment is not limited to what the software does but also found in how the code is written. Sadly, the world is full of things that used to be culturally vital activities, through which people not only found personal fulfillment but earned a decent living, but which are now niche hobbies and only rarely a livelihood, because new technologies provided more efficient ways to achieve the same or similar utility. Calligraphy is an obvious example in the context of this thread. No one is denying that calligraphy is an activity in which artistic excellence and human dignity find expression, or is something that someone could sit down and learn to do — just as they could learn to program —, but today’s civilisation doesn’t need a lot of professional scribes, or for this to be a skill that the majority of literate people need to possess, because we have computers and, before that, typewriters.