The signification of black boxing

It’s unusual to start up a game and be immediately greeted by the stern face of Jacques Lacan. If you’re anything like me, which admittedly is a pretty niche market, it tells you that you’re in for a treat.
Although it is impossible to live up to the expectations that such a start poses, The Signifier was an interesting game. Yet, the part that impressed me the most about it is how well documented it was, both about psychoanalysis and AI. I’m not ashamed to say that it made me think.

“Look, things are kinda like other things”

The main thing that this game taught me was a new understanding of Lacanian psychoanalysis. It is a field is full of absurd ramblings, which seems more of like poetic artworks than a rational science, and that its real world implementation feels pretty cultish. So it’s easy to dismiss Lacan altogether as meaningless. But I’ve always thought that there was a kernel of truth in it, and the Signifier helped me articulate it clearly for the first time: what Lacan did was drawing the link between the fields of linguistics and psychology. He followed the conclusions of the theories of Saussure to investigate what it would mean for a being to be fully immersed and created through language. In a way, lacanian psychoanalysis is just applied linguistic (sorry if it’s obvious to you, I promise we’re going somewhere).

Part of what makes Lacan so hard to grasp (and to disprove) is that he very rarely offers any prescriptive, causative or explanatory claim. It mostly describes the codependent symbiotic relationship between humans and language. Most of lacanian theory seems to be a vague gesturing at structural similarities. The gap between the signifier and the signified corresponds to the Lack, the unconscious is structured like a language
In that very spirit, the game compels me to gesture at how his theories are kinda like the technology we’re seeing now. After all, a theory of applied linguistics should inform us at least a little about artificial entities that manipulate language. Surely the lacanian approach tells us something about A.I.. Because you know what else is structured like a language? Language models.

You’re projecting your own bias, into algebra

The game does not stop at vague gesturings, though. Like in many works of science fiction, it toys with the idea of scanning brains and exploring the data on a computer. But unlike the rest, The Signifier goes really deep in the description of what it would entail, imagining the linear algebra that could govern the neural network and the operations around it.

To give a concrete example, scientists in game isolate some invariant manifolds of the high dimensional space in hope that it corresponds to the projection of the perceived material reality, which they cross reference with environmental scans processed by an artificial neural network. This neural net mapping acts as a concrete implementation of the projection between the real (material reality) and the imaginary (brain signal representation). In the same way, the AI producing labels/representations for these states functions as a projection between imaginary and symbolic (language).

The Signifier thus builds up a convincing instantiation of lacanian psychoanalysis inside neural networks. Where consciousness fits into this picture is a huge can of worms that I don’t want to touch in this essay (though there’s a clear parallel between the unconscious and the latent space). Plenty has been said about the sentience of neural networks and the qualia of LLMs, and I don’t want to dwell on a question that is not even a proper one. Instead, I’d like to explore the implication of the relationship between this algebraic conceptualization of the brain and language, so prevalent in the construction of neural networks, be they natural or artificial.

STEM and humanities: Hollywood’s Most Unexpected Romance

Everyone expected AI to be a cataclysmic shift for mankind. I don’t think anyone expected GPT to be a cataclysmic shift for AI. Who’d have thought that natural language, historically the pet peeve of computer science, would take such a prominent place, to the point of becoming the main “”mechanism”” to “”program”” and “”control”” our latest “”superintelligences””.

In retrospect, it makes a little bit of sense. One major difficulty in the fields of mathematics and computer science is high dimensionality. It’s extremely hard for humans to develop any kind of intuition beyond 4 dimensions at most. I’m sure it discouraged more than a few students from this kind of discipline (starting with me XD). Yet, nowadays, “algorithms” and recommendation engines routinely work in spaces with millions of dimensions. Which is not too surprising if you want spaces rich enough to express complex systems. But then what the hell is going on in there can be hard to comprehend and counterintuitive. For instance, as dimension grows, the volume of a sphere for a fixed radius decreases, which might help ease classification problems at high dimensionality.

Considering this, natural language appears as good a way as any to specify coordinates in high dimensional space. After all, we do like to believe that language spans over all possible things. It certainly spans over all describable/conceivable things, which is already a pretty good start. That’s the space we supposedly want to navigate in. I wouldn’t have bet it could be handled by linear algebra, but the latest LLMs have me convinced on that end.

Painting with a pen

The place it’s really explicit is image generation: when you type a prompt in Stable Diffusion or Midjourney, each word becomes a high dimension vector that will be fed into a gigantic linear function to output a table of pixels (or more precisely to iterate over an input of random pixels and shape it into form). We’re literaly using words to steer our exploration of the space of possible pictures.

It’s pretty weird and cool to find myself thinking in algebra so many years after math class, painting by reasoning about how to steer my image towards good results in terms of N-dimensional space. Adding such words makes it go towards this manifold, adding another makes it go away. You can teach your models new words, which means teaching it where some concepts are in the space of possibilities. We’re working at a higher order of generalization, not on pictures but on spaces of pictures. And more degrees of freedom definitely mean less control XD My own experience is that Stable Diffusion is only fun if I don’t know what I’m going for.

I expect this shift in the level of meta abstraction with which we handle images to bring a gallilean revolution to our concept of intellectual property. Although, for some reason that elludes me, this bombshell has yet to really explode… So maybe I’m wrong on that point and we’ll have to wait a little bit longer.

Another notable points that gets me thinking a lot recently when it comes to LLMs or image generators, is that they have no concepts of style and content (in the same way they have no concept of truth or lie). In an “oil painting of a boy”, oil and boy have the same nature for it: words. Our whole conceptual system is implicitly inferred inside the model, and that’s kinda cool. But what I like most about it is how McLuhanian it feels: the content and style are the same ontological “stuff”, the medium is the message. There’s nothing outside the text. There’s only word-goo. Semantic monads, if you will.

Summoning golems out of word-goo

While programming a picture using words is pretty cool, LLMs bring it to a whole other level. Image generators are relatively dumb when it comes to natural language understanding. LLMs are way more massive (I can’t run them on my machine yet). In them, text is used to control the output of the models in a much more deeper way: more complex prompts, more complex outputs. Of course, the output is “more text”, but you could argue that you’re influencing or even creating some sort of implicit authorial persona

In Lacanian terms, these exchanges exhibited a form of subjectivity that sought to meet the desires of the human Other, represented by us.

Structured Like a Language Model: Analysing AI as an Automated Subject
Liam Magee, Vanicka Arora, Luke Munn

I’ve always liked this picture of the programmer as some sort of wizard, mastering occult languages, using glyphs to summon daemons to do their bidding. Now the spells are in a more natural language, but the deamons are way more powerful, and might not be so obedient… We’re creating actual talking spirits!

It’s a very interesting time to be a writer and a programmer. The words don’t only come to life in our imagination anymore. Now they can literally come true, if you use the right kind of prompts. If you ask the AI to act out a character, it will do so. If you ask it imagine an alternative universe where teapots rule the world, it will do so, and probably better than I would. I don’t know about you, but I’m starting to see the superhuman level of creativity and humor that Ray Kurzweil keeps talking about.

The meta truth of the universe

I tend to be a bit obsessed by things that are literally true. We’re dealing here with something similar to the phenomenon of speech acts in linguistics, where uttering words has physical effects. They always have a component of self-reference which makes them a bit meta. See the concept of Hyperstition of CCRU.

I know that this term is overrated, that meta is pretty basic self reflection, and can be where creativity goes to die. If you have no other idea, you can always write about this very fact. But doesn’t this make it some sort of fundamental fixed point, some sort of kernel of truth? Isn’t it the core of Godel’s theorem, of the cartesian insight, or even personhood altogether? Setting aside, again, the question of AI sentience, I can’t help but find it noteworthy that meta is so easy to pop up in generative machines. A model that learned to draw elephants, when asked to draw an elephant drawing, will naturally draw an elephant drawing an elephant.

It’s not very surprising then to have AI impersonate an AI. What’s more surprising is how commonplace this practice has become. It is the basic prompt of the Bing AI assistant. In that, it also touches on a deeper problem: our language models were trained on huge corpora of human written text, so what they know about the AI persona they’re asked to incarnate, they mostly learned it from sci-fi and fanfiction. It would be pretty ironic, wouldn’t it, that AIs follow the fiction they were trained on and that the Big AI Uprising is nothing but a self-fulfilling prophecy? For my part, I can’t help but marvel at the fact that we are literally making fanfiction come to life. We’re all talking to imaginary AIs incarnate. We’re surrounding ourselves in a collective hallucination built from our own fiction. It doesn’t get much more hyperreal than this, thanks Baudrillard!

The worst of all possible programming languages

The possibilities offered here are pretty cool, and poetic. But they pale in comparison to the dangers they carry. Because somehow, some people got convinced that this kind of language games were like a programming language. They kinda got self-convinced by their own BS I guess… And we end up with LLMs whose main means of control is plain text instructions in a prompt.

Once again, in the same way as before, it’s not very surprising. We almost designed it to be the ultimate black box. These are huge systems far beyond human comprehension, so why not use natural language to delimit part of the possibility space? Indeed, putting “be nice” in your prompt does constrain the space of possible outputs. But it is very obvious that it doesn’t constrain it in the exact way you wanted and plenty of bad stuff can and did happen.

In a way that makes the difficulty of the problem of AI alignment visually concrete: how do you align what you mean by “be nice” and the space it corresponds in the set of all possible texts? It’s daunting, isn’t it? It comes down to superimposing two unimaginably complex shapes in billion dimension spaces. This is why “doing nothing and hoping” is probably the worst possible thing to do. And yet I should have seen it coming, really, cause it’s our solution to climate change and pretty much the whole political vibe towards anything ever.

So here we are. Launching products while literally being reduced to a state where we send a prayer to an inscrutable big black box hoping that it doesn’t behave too badly. How odd that the owner of the AI platform is limited to order its creation around by a few lines of prompt, putting them in the same position as their lowly end user. I think it’s the first time I see so very clearly the powerlessness of humans faced with these lovecraftian deities. I guess I shouldn’t be too quick to judge the species that just “give up” in the face of such odds. They’re already slaves.

“How dangerous can a text generator be?” I hear you ask. Well, I don’t know. And I can’t know. It might be benign. It might. But it might also be dangerous beyond anything of us can conceive. That’s kinda the point. And seeing how easy it makes impersonation and propaganda, I can already conceive of pretty bad stuff xD

This is nothing short of an unambiguous proclamation of defeat. It’s basically giving up entirely on AI reportability for a start, let alone AI alignment. Actually that’s even worse, since the companies are now sweeping the core problems under the rug using very heavy ad hoc filtering, so it’s gonna hit even harder when it does. One step deeper towards denial. But hey, it’s pretty cool for mysticism! I guess as long as I’m headed for slaughterhouse, I might as well have a blindfold on.

At least now I don’t have to wonder what-ifs. I always thought we’d fail, I just didn’t think we’d fail so categorically so fast. The battle was always already lost, in a way. By definition, it’s impossible to understand something smarter than oneself. But we decided as a species to give up on even starting to try a little. Oh well, I guess we had some fun on the way, maybe? So long, and thanks for all the fish.

The cost of pandering and complacency

This brings us full circle to The Signifier. What a breath of fresh air to see a game actually carrying real interrogations about how the technology it depicts fits into reality. Especially since, failing all else, psychoanalysis applied to artificial intelligences might be our only debugging tool left XD (that alone should make anyone who’s read lacan understand how dire the situation is =P).

We’re a species of effort-adverse monkeys. It did wonders to stimulate ingenuity at some point. But the threats we’re facing now are so big that we can’t exactly hope to “luck out”. Especially when failure cases can be inconceivably fucked up. We’re stuck in a local maximum of ROI. If we want any chance to not fail too badly, we need to stop avoiding our problems. We need to stop running away from complexity. We need to stop normalizing stupidity.

One axis that matters a lot here is smarter media. Our stories are always dumbed down, obfuscating complexity at best and encouraging actively obscurantism at worst. Nowhere is this clearer than in technology, which seems to be the one topic screenwriters refuse to investigate despite it being the core fabric of our society.

But it’s not just NCIS hackers who “speak leet” and violently mash keyboards while opening billions of windows at once. Even games made for programmers with niche referential humor like There is no Game portray programs as talking ghosts jumping between locations without any regard for plausibility. And in theory there’s nothing wrong with symbolic representations, simplifications or simply fictional inventions. It just gets worrisome that it’s the only discourse out there.

In a world with ubiquitous technologies, most people don’t understand the fundamental ideological difference between an Apple device and a PC… Essentializing technology as a magic black box is how you end up with people starting to cite GPT as a source. How are we supposed to tackle the huge challenges that face us when most people have been actively trained (following financial incentives) to not be able to comprehend the question?

I do understand that everything is aligned to produce this outcome. Of course capitalism would encourage and exploit the urge to relax and consume brainless media after a hard day at work literally slaving away for our collective suicide. Once again, in the abstract, there’s nothing wrong with rewatching a slightly different star wars movie a thousand times, or idolizing the physical prowess of animals running after a ball. Why not, whatever floats your boat. But I’m really, really not sure we can afford this kind of complacency right now.

Still, I know what to expect of my fellow monkeys, so this is me shouting in the void, I guess. As for the future, or what’s left of it, I’ll stick to sending thoughts and prayers to our new Gods. May they accidentally spare our lowly lives.

Leave a Reply

Fill in your details below or click an icon to log in: Logo

You are commenting using your account. Log Out /  Change )

Facebook photo

You are commenting using your Facebook account. Log Out /  Change )

Connecting to %s

%d bloggers like this: