“…It feels strange. In a way […] you would think it would be strange to do it every day, but you get this kind of feeling that it feels strange actually just the first time you do it. The second time it’s still exciting and the third time it becomes work. Because you have to keep doing it over and over again…”*
*on breaking glasses for machine learning algorithms
Steyerl, Hito. The City Of Broken Windows, 2018.
-Do you know what turns darkness into light?
-Poetry.
Godard, Jean-Luc. Alphaville: A Strange Adventure of Lemmy Caution, 1965.
If the sum of human activity can be polarized between two extremities, some language of heavens that aspires to finally be able to show everything and some language of earth that endeavors to finally be able to know everything, then the present condition can be described as a garden of machinic, bitwise, automated delights. Algorithmically controlled gathering, processing and creation of information, via machine learning applications, accumulates to a shift of focus: rather than understanding or creating artifacts inhabiting reality, we wishfully create systems that will hopefully understand and create those artifacts instead.
This infiltration of automation, accompanied by its mandatory critique, has a long lineage of ups and downs, springs and winters as the customized terminology would dictate. Historically any progress made within the industrialized field of automated knowledge/creativity, is followed up by some withdrawal partially attributed to socioeconomic factors or material limitations but principally caused by a skewed expectation versus reality feedback. When it comes to artificially generated content, added to evident expected surplus values (speed, quantity, efficiency and responsiveness to list a few) and prominent among them, is the one of automation “passing”1 as craftsmanship. This passing, enhanced and deformed, acquires an ontological state of discrimination between fake and true, where fakeness is not to be merely perceived as forgery of truth but rather as a dull, emptying out of it while its negation, truth, remains to be unraveled.
Whereas the old saying wants the magician to be an actor playing the part of magician, there is a strong connection binding them with their tricks, not just in terms of competition -after all the performative aspect of magic is equally important with is guidelines- but mostly because when magic is revealed, any possibility of performance under its mesmerizing cloak turns obsolete. To be able to evaluate cognition, one must be able to define and recognize it. But to be able to define what automated cognition is expected to be, one must be able to demarcate what general cognition is in reality from what general cognition does to reality. This confused differentiation between identity and appearance, what something is versus what something does, results in a tasked approach that regardless of seeming quite vague at higher levels of abstraction (a generic term such as cognition), unfortunately becomes more prevalent at the lower ones (particular terms such as pattern recognition, content generation and so on). Every estimation of the aforementioned passing value, true or fake, must be traceable within the context of the definition of its grammatically connected noun. However it remains idiomatic of artificially simulated intelligence and creativity, that every time a piece of this entangled landscape of deconstructed task boxes gets figured out in truth terms (real results), the resulting response is that it is mere computation (fake means).2
Thus, a neurotic relationship with the tool emerges: on one hand the creation of an omnipotent, omniscient and omnipresent generative system appears to be the holy grail of creation. On the other, its potentiality is already predetermined by a conformistic adherence to an already known standardization. The goal is never to be achieved, because the goal is always what has not been achieved yet. Valleys and mountains of two-dimensional curves, that estimate the progression of this neurosis, reveal that the more something made to resemble subjectivity starts to really resemble subjectivity, the more uncomfortable it gets.3
In the meantime computational generation, with the verdict of its authenticity still pending, has already permeated the stronghold of the real. Its applications subtly echo through the whole spectrum of sensorial and mental culture. Interconnected digital antennae broadcast -online and offline- an algorithmically produced signal, of connotationally aligned datasets, labels and relationships, allowing their receivers (users, co-authors, laborers, clients or products) to consume them, follow them or pin them, in a quasi-cannibalistic manner, where tomorrow’s meal gets produced from the leftovers of yesterday. This thread of amplified automation, both functional and ideological, subverts in its turn real world signification or meaning into a mere array of symbols -timelessly repeating regardless the question and thus ultimately indecipherable- casting the physical, human reciters bot-like figures,4 while the captivated audiences still wonder if the old trick of turning them off and on again will do. At the same time everyone checks the box of “No, I Am Not a Robot” dictated by security controls of the constellation, subscribing and negating simultaneously the stigma of mechanic repetition of creation, because the more they answer it the more unsure they get.5
Of course automation, oscillating within a conceptual debate of efficiency and inefficiency, easiness and difficulty, authenticity and replication, comes with its own limitations. So far, its most important achievements stem from a weighted function that allows it to predict subsequents from the antecedent. Its foundation to statistical induction,6 where given enough information in the form of primitive data a system can construct some form of logic preoccupying their arrangement, does not prevent the hope that an alchemic formulation of sorts is feasible and will inevitably reveal behind them, those things that connect them, initiate or consummate them, a transmutation from tiled instances of binary nature to a longed-for meaning, the bigger picture behind the dots.
However prediction -projecting in the future the knowledge of yesterday- differs considerably from creation -casting a difference inside repetition. The task of poetics usually favors possibility -or even the impossible- against predictability. “If he [the artist] describes the impossible, he is guilty of an error; but the error may be justified, if the end of the art be thereby attained” the philosopher will note.7 And a statistical inference of the new -that is not merely combinational but exploratory or transformational instead8 – would indicate a blissful forgetting of what’s learned -the right of being out of given truth. For even in the most relaxed examples, the learned curve must indubitably fit inside the data; it must describe their categorical ontology, with maximum efficiency and minimum error. But creation is still that, which has not been done yet and if a nonsense arguments like “if art is made of green chips, the world is coming to an end” might never stop messing up with logic, as paradoxes of material implication, they can still count as a starting point, of other things to come, regardless.
After all, prediction relies on compression. Noise is an unwanted anomaly scrounging off the clarity of data, responsible for all sorts of difficulties and inefficiencies in the context of automated learning. The reason why “These Cats Do Not Exist”9 is not to be found in terms of passing. After all they do appear real enough, all things considered. Their inexistence is to be examined in the abstracted attributes that affect their instantiation10 and in-between their imprecision or incorrectness -wrongly assigned attributes- there still remains plenty of room for the unobservable -additional attributes, forgotten. This latent space, the representation of compressed data -quantities that are not directly observed but rather inferred from other observations11 and can conclude to qualitative difference- will eventually be called to stand for a cat, a person or an artifact. It can be perceived as a consequence of the very notion of automation and is organically bound to the very architecture of generative, machine learning algorithms. In these compressed data symmetries, law-like predicates such as visual perspective or point of view, texturing and interface boundaries (the contact surfaces between objects that stitch a web of time-space reality), are not explicitly declared but rather assumed from arrangements of grid colorization (pixels on screens). This compression without orientation results in a yet to be understood absurdist tendency, where the reward is being able to “name at least one thing in this photo”.12
If a generative model is to be understood as its cybernetic equivalent, with input-output feedback loops and calibrations in-between, it can be perceived as a graph-system of nodes that correspond to variables and directed arcs between them that represent dependencies, causalities of sorts.13 The top-down view of this system is supposed to coincide with a generative process of content creation; however it remains an executive synecdoche, where many crucial parts of the process (e.g. creator, point of view, medium, history, time and matter, sensing variables) remain hidden or irrelevant, what gets produced is what will matter. Reminiscent of an incompleteness theorem, the question of creativity seems unanswerable to models replicating it, because it is not to be found in the graph, not even in the whole model simulating it, however complete and consistent. On the contrary, it is to be located in the paradigm shifting, the very building of the model, a worldmaking and from there to a construction of reality, a newness that corresponds to hows and whys rather than whats.
Expecting to approach this problem -a misleading term, to begin with, since creativity is rarely perceived as a problem in need of solution, but rather as a possible solution to problems- in terms of turning subjectivity into objectivity, by providing ever more content, ever more resources, ever more abstraction, results to a skewed view, never out of the model but -oddly enough- from as within it as it gets. The statistical file generator is trapped in a closed universe predicament, whereas expected objectivity never arrives, but its summed up, weighted subjectivities –human, all too human– become reality regardless. The hope that this excessive, periphrastic, descriptive methodology in generative context -an unlimited semiosis of sorts- can still produce a difference, remains of great value to the author of this short essay. However, it entails the danger of becoming a deluged representation of reality, matching the extents of a data empire itself, that does not manage to forestall the occasion, where even when a monkey hitting the keys of the machine in random ways results in Shakespeare, a choice is still to be made between the slim light of words and the vast darkness of nonsense.
Can art be made of green chips? The question is to be perceived primarily as an interrogation of human activity. If systems seize to be anthropomorphized, be projected upon as helpless collaborators from one side or as dreadful competitors on the other, a new option does emerge where artificial creativity gets to be a mixed science and art of imaginary solutions. The limitation of definitions, fixed in this moment of becoming, amplified by unknown views and restricted by the already established, unfolds as a new language. Thus, it becomes a speculative framework, flirting with the state of what is possible. It is a turning point that automation stops to function, yet in a schizoid way still performs by revealing the ubiquitous complexity behind tools, methodologies, prima materia, and in the end intentions. The CAPTCHA test that concerns us, then, is not a misinterpreted evaluation practice, oversimplified and ignorant of its initial value, but one that highlights “the differentiation between music sheets where if you are to add a decrescendo it would add to the musicality of the piece without being interpreted as an overly heavy-handed metaphor within the context of the thematic material”.14 Impossible decisions of sorts highlight facts different of anything observed15 and bring new data to the table.
After all, even if the destination seems unreachable, the question embarks on an interesting journey. And one cannot help but recall a rather popularized version of it: in the movie “I Robot”16 where the protagonist, faced against a superintelligent droid, asks if it can write a symphony or turn an empty canvas into a beautiful masterpiece, only to be faced with the most genuine counter question: “Can you?”
_____
References
1. Bratton, Benjamin H. “Outing Artificial Intelligence. Reckoning with Turing Tests.” In: Alleys of Your Mind. Augmented Intelligence and Its Traumas, edited by Matteo Pasquinelli, 69–80. Lüneburg: meson press, 2015. DOI: https://doi.org/10.25969/mediarep/1282.
2. Indicative of this statement is the prominent AI effect: “It’s part of the history of the field of artificial intelligence that every time somebody figured out how to make a computer do something—play good checkers, solve simple but relatively informal problems—there was a chorus of critics to say, ‘that’s not thinking’.”- https://en.wikipedia.org/wiki/AI_effect
3. Mori, Masahiro, Karl F. MacDorman, and Norri Kageki. “The Uncanny Valley.” IEEE Robotics & Automation Magazine 19, no. 2 (2012): 98–100. https://doi.org/10.1109/MRA.2012.2192811.
4. Ono888. “These Strikes Are Wrong – Bot Ed Miliband,” Youtube Video, 2011. https://www.youtube.com/watch?v=wCem9EZb-YA .
5. Bratton, Benjamin H. “The New Normal, Hemispherical Stacks, and Algorithmic Geopolitics”, Lecture at Romantso, Athens, Greece, 5 September, 2018.
6. Pasquinelli, Matteo. “Machines That Morph Logic: Neural Networks and the Distorted Automation of Intelligence as Statistical Inference.” Glass Bead 1, no. 1 (2017).
7. Butcher, S.H., ed. The Poetics of Aristotle. Third. London: Macmillan & Co, 1902.
8. Boden, Margaret A. The Creative Mind, Myths and Mechanisms. Second. London: Routledge, 2004.
9. https://thesecatsdonotexist.com/ – Generating Cats with StyleGAN on AWS SageMaker
10. Alpaydin, Ethem. Introduction to Machine Learning. Cambridge: The MIT Press, 2004.
11. By definition from Wikipedia: https://en.wikipedia.org/wiki/Latent_variable
12. A famous meme calling its viewer to discern entities in an image, that looks normal at first glance but after close inspection is indecipherable, since all the entities melt together into an indistinctive amalgam. https://knowyourmeme.com/memes/name-one-thing-in-this-photo
13. Alpaydin, Ethem. Machine Learning. Cambridge: The MIT Press, 2016.
14. A memefied version of the regular CAPTCHA test, where in order to prove they are not a robot, users have to select images of music sheets with the aforementioned qualities. Available at https://music.stackexchange.com/questions/114012/what-does-this-music-captcha-mean
15. Peirce, Charles S. “Deduction, Induction, and Hypothesis.” Popular Science Monthly 13 (1878).
16. Proyas, Alex. I, Robot, 2004.