Mario Carpo

Publication History:

“Rice or Pasta? Choose your AI,” in Diffusions in Architecture. Artificial Intelligence and Image Generators, edited by Matias del Campo, 2-6. Hoboken, NJ: Wiley, 2024

The text posted here is a preprint draft and it may be different from the published version. Please only cite from copy in print.

Many years ago, on a warm, early summer day toward the end of the last millennium, I was visiting my parents in my ancestral hometown at the feet of the Western Alps, in Northern Italy.   I was traveling in the company of my then girlfriend, a born and bred New Englander, who was of course much intrigued by the traditions and customs of that old, primitive land.  I watched your mother cook risotto last night, she told me: and I am certain she never used any measuring tool.  She has no scales in her kitchen; nor measuring cups for flour, grains or liquid.  I did not see a thermometer, not even a dial on the oven; nor a timer; and I am pretty certain she never looked at her watch.  How can she cook?  I have no clue, I replied; evidently, the risotto was edible; why do you not query her, discreetly, tomorrow?  And so she did, and it turned out that my mother’s reply to every query involving the use of measurable quantities in her cooking (how much of it? to what temperature? for how long…?) was “well, it shows”; or, “you can tell” (apparently, my friend inferred, my mother’s telling was based on color, odor, or touch).  My friend concluded, while we were on our way back to the airport a few days later: your mother—and, I think she added, this entire country—seem to live in timeless universe of approximation; the modern world of precision, and the spirit of modern science and technology, based on numbers and quantification, never made it to these shores. 

I think that around that time we had just finished our seminar readings on Alexandre Koyré, so we may be forgiven for having momentarily forgotten in that context that the scientific revolution itself was born not far from the small subalpine town where that memorable conversation took place, regardless of Galileo’s diet (on which, by the way, much is known).  Also in retrospect (but we didn’t know it back then), in our modernist critique we were somehow attributing to my mother many of the traits and features that a bit later in time came to define the post-modern, “phenomenological” craftsman, as idealized by Richard Sennet and others ideologues of that ilk, who see the artisan as a man without words and reason, who makes things by feeling and intuition, incapable of explaining to himself—or others—the logic of his doings.  All artisan lore is thus seen as “tacit knowledge”, an embodied form of technical expertise which can be neither learnt nor taught, if not by some form of mystical empathy (Einfühlung) between sentient souls and more or less animated objects. 

In so far as my mother’s case is concerned, I suspect we might debunk some of these theories if we could assemble a team of scientists, comprising chemists, physicians, biologists etc., and ask them to observe her cooking over a consistent duration of time, collecting then averaging and converting their quantitative observations into formal rules that anyone could then follow at will.  However, as it happens, we do not need to do that—nor does it appear that anyone ever did—because modern societies have found an easier way to bypass my mother’s kind of artisan knowledge.  My mother’s skills in this anecdote were due to, and inherent in, the food she mostly and primarily processes—rice.  Locally grown rice is a natural product, retailed even after milling and polishing pretty much in its natural state—with grains in different sizes and shapes, each with variable, unpredictable contents of starch, to be released during cooking in equally unpredictable ways, which only an experienced cook can detect, interpret and adjust to as needed. 

If you do not have the talent and patience to do that, do as I do—cook dry pasta.  Dry pasta is an industrial, factory-made product; its prime component is a blend of different wheats, carefully balanced to obtain steady chemical and physical properties; after mixing with water and other ingredients the dough is steamed, cooked, and dried twice, before and after extrusion and cutting, in controlled conditions of temperature and humidity.  In the end, the pasta we get from the grocery is a standardized product—as standard as a bottle of Coca-Cola or a hot-rolled I-beam made of structural steel.  This is why its cooking time in a pot of boiling water can be predicted, always the same and the same for all—so long, that is, that the water in which we boil it contains the right amount of salt; in case of decreasing atmospheric pressure one may need to recalculate the cooking time using a formula which I have honed during a life of mountaineering and can forward for free to interested parties on demand.  The point is, the entire process can be formalized (converted into rules) and scripted; and when it is scripted, it can be carried out by any person without skill or prior training in the subject—or by any suitable industrial robot, duly programmed.

Artificial intelligence was always meant to do way more than that.  In its original formulation (see for example Marvin Minsky’s seminal Steps Toward an Artificial Intelligence, 1960-61) artificial intelligence was seen as a “general problem-solving machine” that would find solutions following iterative trial-and-error routines (using strategic shortcuts whenever possible: gradient-based optimization was the smartest and is still largely in use today).  In so far as the machine was expected not to repeat the same mistakes, i.e. to learn from its own mistakes, this was not unlike what today we call machine learning—even if the term itself came later.  However, when it became clear that the machines of the time were not powerful enough to work that way, this strategy was abandoned and soon replaced by a more practical one, known to this day as “knowledge-based”, or “rule-based” artificial intelligence.  In this new mode the machine is not supposed to learn anything at all; instead, a selected corpus of human knowledge is instilled and installed in its memory, conveniently translated into a set or arborescence of formalized rules, which the machine will implement step by step when prompted.  Architectural historians may remember, as an instance of this logic, Negroponte’s 1970 Architecture Machine, which was meant to lead users to design a house via a sequence of if/then queries and multiple-choice options.  Human users of the machine were not required to have any knowledge of architecture, because all architectural knowledge deemed necessary to solve that specific set of design problems had been pre-installed in the machine itself.  That particular “expert system” famously never worked, but many later and less ambitious ones did and still do.  To go back to my anecdote, any such machine, when paired with adequate sensors and mechanical extensions, would be perfectly capable of cooking a perfect serving of pasta.  The revival of the original mode of artificial intelligence (now often capitalized and abbreviated as AI) started some 10 or 15 years ago, when it appeared that due to the unprecedented computational power of today’s machines (Brute Force computing) and to mere size of today’s searchable memories (Big Data), today’s computers can indeed solve many unwieldy problems if they simply keep trying.  This is how a relatively stupid but massive trial and error strategy came to be seen as key to delivering a functioning artificial intelligence, and indeed the almost miraculous successes of today’s AI are due for the most part to machine learning, not to rule-based computation.

The design professions had their first encounters with the surprising image-making potentials of this new breed of born-again AI following the invention of Generative Adversarial Networks in the mid 2010s.  As we now know full well, GAN can detect and analyze some commonalities inherent in a consistent visual dataset (a corpus of selected images), extract them, and convert them into a vectorial definition (known as “latent space” in technical parlance).  This mathematical abstraction can in turn generate new images similar to those in the original corpus; or can be tasked to combine features taken from that corpus with new images, thus creating hybrids between datasets where, in particular, some formal attributed extracted from one corpus are applied to the content of another.  Not surprisingly, this operation was called “style transfer” by the engineers who first described it in 2016;[1] one must also come to the almost inevitable conclusion that GAN can carry out, in this instance, something similar to an intelligent inductive operation which, by dint of comparison, selection, generalization, abstraction, and formalization, automates visual imitation—a cognitive process which has fascinated and eluded philosophers, artists, and scientists since the time of Plato and Aristotle. 

The tsunami of Generative AI that has swept the world since the spring of 2022 has not significantly altered this conceptual framework; in fact, for visual artists and designers, the text-to-image stunts of DALL-E etc. are not particularly interesting—at least, not for the time being (even if they are admittedly disturbing, or worse, for linguists and writers).  A research paper published in the fall of 2022 by Google’s DeepMind,[2] which promises to extend the ambit of imitation learning beyond the ambit of styles, pictorial or otherwise, and apply it to knowledge transfer as a whole, may likewise not be of immediate consequence for the visual arts, but it may significantly impact the art of cooking, and—at the opposite end of the hierarchy of intellectual added value—the modern theory of science in its entirety.  The DeepMind team, noted for its expertise in imitation learning, which they famously nurtured and honed by teaching computers to learn games, has now trained a robot to learn how to sort a set of colored blocks by observation alone—i.e., by feeding the system a huge dataset showing examples of said sorting, which the computer learned to replicate (i.e., imitate) in the absence of any formal rule telling the system what to do, or how to do it.

This machine, if fed a sufficiently vast dataset—for example, given the possibility to observe my mother’s cooking over a lifetime—could learn to replicate my mother’s risotto-making skills by imitation alone.  It would do so by assimilating, then replicating, the unknown and arguably unknowable rules underpinning my mother’s cooking; in true AI style, it would do so without ever spelling out said rules—as, once translated into the system’s “latent space” (a multi-dimensional vectorial space), these cooking rules would be perfectly readable, but utterly meaningless to any human intelligence.  Since tacit knowledge, as conceptualized by old-school phenomenologists, is mysterious by definition, this mode of AI could then, simply, automate tacit knowledge—sans its magic aura and hocus-pocus. 

Even more than the applications, the implications of this mode of machine learning are staggering.  Whenever a functional (i.e., big enough) exemplar corpus can be gathered and marshaled for training purposes, we must assume that imitation may replace traditional rule-based procedures.  So for example, an engineer given a structural model of proven functionality but analytically incalculable, may decide to tweak it to adapt it to a different context by imitation alone (so long as the AI system may map each transformation of said model to a “latent space” in turn derived from a sufficiently vast, relevant dataset).  This is not unlike what modern science always did—after collecting a corpus of experimental data, sifting, comparing, and generalizing their results; then formalizing rules that can in turn be applied to predict the unfolding of unprecedented but similar events.  The only difference is that, this time, we won’t do the job, because the machine will do it; and we won’t know the rules, because the machine won’t tell us.  So, in a nutshell, here is the tagline for tomorrow’s engineers: when you cannot calculate, imitate. 

Is this not what craftsmen always did?  This is what computation today does best.  But, let’s not beat about the bush: this also means the end of modern science—that which came out the Aristotelian, Scholastic, Galilean tradition; the modern science of which computer science itself is the latest avatar to date.  To be precise: this does not mean, literally, the end of that science itself—as we would still need it for many intellectual reasons.  It means the end of many of its practical functions.  But it’s a slippery slope, as some alternative sciences are, as always, lurking in the dark and are ready to jump onto a weakened prey.  There is, of course, a viable alternative to all this—toward which I incline.  Drop the risotto; stick to pasta. 

[1] Leon A. Gatys, Alexander S. Ecker, and Matthias Bethge, “A Neural Algorithm of Artistic Style,” ArXiv: 1508.06576 (August–September 2015); Leon A. Gatys, Alexander S. Ecker, and Matthias Bethge, “Image Style Transfer Using Convolutional Neural Networks,” 2016 IEEE Conference on Computer Vision and Pattern Recognition (June 2016): 2414–2423.

[2] Scott Reed, et al., “A Generalist Agent,” Transactions on Machine Learning Research, 11/2022, arXiv:2205.06175