# Text edition Page: https://pleometric.net/archive/sadly-im-going-insane/ ← Archive # 2ch on AI and artistic labor This text was used as the source script for the video [Sadly I’m Going Insane](/videos/sadly-im-going-insane/). The script is taken from a bitter rant originally posted on the Russian imageboard [2ch](https://web.archive.org/web/20220814132912/https://2ch.hk/pa/res/787542.html#809450) about AI, artistic labor, and the obsolescence of creative institutions, translated by [@teortaxesTex](https://x.com/teortaxesTex). English translation transcribed from four supplied screenshots. The overlapping paragraph is included once. The screenshot identifies post 809450 and displays “17/09/22 Sat 17:27:32”.
Yap-yap. "A neural network is just a tool." Instrumentality is not an essential characteristic, not a totality of immutable and stable invariants — it's merely a position in a network of relations. Photoshop is your tool. You are a tool in the hands of the client. The client is a tool in the hands of media corporations that have flooded the information space with hyperstimuli, pussy-triggers, freak-codes, Miyazaki-glitches, and other visual-advertising slop. In this hierarchy of instruments, you're merely a gasket. You're tolerated only for the time being, used as a cash cow for creating datasets. While corporations keep a death grip on their "know-how," they sell "sharing is caring" to the goyim from every loudspeaker. So, how's the sharing going, boys? Like it? Want more, or has it finally sunk in? > > 809301\ > > Fortunately, there's me, who understands both neural networks and drawing at the same time Judging by the millet you've been chewing here, you don't understand either. What's worse, you don't even understand yourself, and your doctrine of the human being is hundreds of years out of date. You see, the future is distributed very unevenly across humanity. The conditional present is an integral sum of anachronisms; the civilizational clocks within different social strata either lag or rush by whole millennia, and while fiery sparks of futurism soar skyward, somewhere deep beneath the silvery amalgam of modernity, tectonic strata of the archaic heavily stir, with roots reaching straight into the Paleolithic. I've seen \[fields, Neo, endless fields\] entire institutes filled with former humans incapable of recognizing the moral obsolescence of their own discipline. They live, and sometimes even live quite well — but their life relates to modernity the same way the life of a bushman poking the ground with a stick does. You are a former human, a representative of the Paleolithic. You massively overestimate both your own complexity and the complexity of the culture that spawned you. You menacingly wave your poking-stick and seriously believe that your so-called "academic drawing" is something more than just a morbid art of projecting probability distributions learned by the visual cortex into a two-dimensional peephole. Glory to the great academy! Glory to the dung ball molded by the calloused hands of industrious dung beetles — a totality of peasant intuitions, a-bit-here-a-bit-theres, and rules of thumb, lovingly passed from leaky teachers to equally perforated students! It doesn't even occur to you that instead of "constructing a nose" as a perspective projection from 3D to 2D — which some three-dimensional little humans spend years forcing upon other ones — you could "construct a face" as a generalized object in a 1,000,000-dimensional space — and obtain all potentially possible human faces (i.e., those fitting within the probability distributions learned by the model) as projections of that generalized object onto a plane. The neural network doesn't need to "construct a nose" — it's already constructed within it as an element of a certain hyperobject in hyperspace — and, unlike you, you fool, it can project this nose into a subspace of any dimensionality: into a line, into a plane, into a three-dimensional volume, into a ten-dimensional volume, anywhere. "Construction" is not your feature — it's your bottleneck. The mandatory decomposition of an image into hundreds of Photoshop layers, primings and underpaintings — that's yet another bottleneck, a crooked and deaf detour where the neural network can cut straight through. Local single-channel application of an image using a hand-manipulator, the slow and semi-random coverage of a surface with monotonous brushstroke patterns, all this stuffy, mothball-reeking, chthonic horror — it's simply creative disability compared to a neural network that builds the image globally and multi-threaded across the entire canvas surface, and which, unlike you, is equally capable of building objects of any dimensionality in a global and multi-threaded manner. > The real point is that artists actually get paid precisely in those areas where this very intentionality is needed. So that a concept of a girl for a mobile wank-game has every little string on her backpack and pocket on her jeans there for a reason. Ah, there it is — pockets on jeans. A veiled attempt to appeal to the non-formalizability of pragmatics. The weakest argument for props I've ever encountered. By the way, do you remember what year Ackoff's "On Purposeful Systems" came out? Seventy-two, was it? Locked in the sealed crypt of their own stupidity and ignorance, representatives of the creative class still consider even long-solved problems to be cognitively impenetrable. > So that she'd be redrawn forty times from different angles and every little detail invented for her — how she expresses emotions, how she's built, how she moves. A neural network would just fart out a bimbo toilet with random smears instead of details, and that's where it all ends. Ten million "intentional" crooked and deaf detours, thousands of tons of glucose burned in the cracked pots of the creative class — and the output is still the same ten standard plots and a hundred or so classical archetypes. That's how functional-structural convergence works. That's how power law distribution works. That's how the ethical and cognitive specialization of sapiens works. A finite number of basic emotions, a finite number of their variations, a finite number of forms of group interaction, a finite number of social conflicts, a finite number of characters in the actor network — "love" this way, "betrayal" that way, "family" to the right, "war" to the left. Locally, you're all incredibly individual and monstrously creative — but the moment you step out onto the expanses of ArtStation, or even the most obscure booru, you want to grab your head. Tits-woman, dick-woman, strong independent woman — the exotic maiden of the eggplant variety. What isn't observable on small samples is perfectly visible on big data. It's all the same shit. ------------------------------------------------------------------------ Damn, this is hilarious as fuck. "Artist development." "Income." What income now? The base model understands Euclidean 3D in classical perspective. Why this particular space? Because this abstract model most efficiently explains the presented totality of data. Why does it understand "objects" and the group of orthogonal transformations? Because it's the most convenient way to compress data through hierarchical decomposition of the visual field. Why does it understand lighting? Because, just as in previous cases, the model managed to reverse-engineer the scene and decompose it into surfaces, light sources, and color space. Why does it understand "style"? Because "style" is a microscopically small, in terms of information volume, cosmetic overlay on top of the base geometry, textures, and lighting. And yes, "composition" is also an element of style. As are various types of perspective. For a manga artist, "fish-eye" means years of struggle against built-in visual cognitive biases, painstaking fiddling in an uncomfortable 2D, isolating, combining and projecting primitives, generating and tracking pencil trajectories, an iterative cycle of rote-memorized algorithms with elements of pseudo-randomness — assembling a cognitive chimera from controllable elements of thought. For the model, it's merely a vector weighing a few kilobytes, controlling hierarchical blending of base elements. Shift by a vector plus cosmetic rendering. Click — an orange sprouted feathers. Click — the feathers turned to glass. Click — fish eye. Click — reverse perspective. Click — Kuvshinov. Click-click — Vrubel. Gumroads never actually taught anyone anything. Creativity is a pipeline. 99% technical tricks — and a little left to chance. And this isn't even the beginning — it's merely the anteroom. Text, images, music, video, speech. Sculpture, architecture, engineering. What in our culture is truly difficult? Mathematics? But theorem proofs are isomorphic to programs — and coding, well, is also kinda, you know. For the market, only meat and qualia will remain. You can't bring qualia to market. Therefore, only meat. I'm just complaining. We've all been blacklisted long ago into a state of petrification. The point is something else. Creativity is now officially a closed topic. You know, quite recently, among linguistically oriented highly intellectual people (philologists, writers, philosophers of language, mathematicians...) it was fashionable to do topology, play word golf, admire Hofstadter, and subscribe to the magazine "Word Ways." From those glorious times, only scraps, shards, and rudiments remain in the form of ubiquitous crossword puzzles and scanwords. The end of the glorious era came almost instantaneously by historical standards. And not because fashion changed and people got bored with something. It's just that a cheap microchip appeared, capable of closing all open problems and finding all interesting structures in about an hour — even without any algorithms, just brute force, by enumeration. The topic was closed. And nobody even noticed how it happened. It happened too fast, almost immediately. Flip-flop, and that's it. Poets spent centuries constructing palindromes. The microchip came and found them all. Aristocrats spent decades playing word ladders. The microchip came — and found them all. In the thickets of texts, there once roamed the most amusing Pokémon — but the microchip came and caught them all. Dickgirls, dickgirls — dickgirls, vegetables... Painful, not painful — is that even the point now? Ceilings no longer decide anything. There are no ceilings. That's it. Flip-flop on the scale of an entire culture. I didn't manage to slip through in time. And there was so little left. Ah well. ------------------------------------------------------------------------ > Well, people decided to make a little project, +1 tool to an already long list. Fractal laughter. Chuckles within chuckles. The point is multimodality — the practically realized ability to reconstruct a latent model of the "world" from two qualitatively different discrete projections of it. Instead of natural language (or rather, the truncated version used for tagging) and a chunk of N^N characterizing geometric and color space, you can take any two models of anything, recorded in any two arbitrarily chosen languages — and build from them a proxy fantasy-world that guarantees coherent inference in each half of this bicameral. This is the same "Geometric Machine," only built on self-learning, giving x1000 in power, and additionally capable of running on any cardboard placeholder rather than only on a distant mainframe for the highbrowed. Multimodality allows you to bypass local dead-ends by switching representation, enabling you to go around the problematic area along a parallel model — either by ascending to a different level of abstraction, or by decomposing the problem into subtasks within a more concrete and detailed representation. By linking a hundred bicamerals on different pairs of models into a directed graph, we get the very thing the fathers of cybernetics dreamed of, and which fifty years ago the blessed Academician Ivakhnenko wrote about in plain text: a multilayered hypersurface of model-layers, connected by forward and backward links of information archiving and inflation. Considering the fact that such structures of models docked into polyads keep surfacing in the most diverse fields related to thinking — from mathematics to cognitive sciences, from Lotman's cultural studies writings to articles on pedagogy — one can hypothesize that they are universal to the same degree as the g-factor, cortical columns, deep learning models, and other abstract gnats like atomic theory in physics, set theory in mathematics, or the finite element method in engineering. In any local domain, sooner or later a universal methodological wunderwaffe is discovered that completely covers the problem area. Understanding is the transformation of information within a network of models representing other models and themselves — a long branching tail-cyclic graph, a rhizome of algorithmic tunnels propagating a long echo of mutual interpretations through the system, a reverberation of meanings in self-referential algorithmic corridors. Did you understand what I said? Neither did I. But that's how it is. We perform this transformation every day, every second, within hundreds of parallel threads — and we use its results in the form of object icons on a screen and a dreary linguistic construction set, with which we desperately try to combat reality. Taking jabs at Vasya won't help you. The issue isn't about riffraff — it's that "meaning" has become an engineering problem, literally just now, right before your eyes. And instead of assessing the existential risks and possibilities for hedging them, you grabbed a couple of dumb verbal parts and are trying to assemble from them either a butt plug or a virtual inner prison guard. Repent, simpleton, you'll get a discount.
[Watch Sadly I'm going Insane →](/videos/sadly-im-going-insane/)