On Large Language Models, aesthetic sense as knowledge, and the material limits to “artificial intelligence"

via whale language, living consciousness, and the poetry of Aimé Césaire

An interesting week in “AI”: ChatGPT solved an 80-year geometry problem after responding to a regular chat query, a prestigious literary committee selected fairly obvious LLM slop, and in one of my classes, 100% of my students not only submitted LLM essays, but then sent LLM follow-up emails asking me to regrade them. (I have a feeling the confidence requisite to pull such a stunt is also encouraged by GPT etc.)

Superficially, it seems that AI is a mess of contradictions, a “genius” in some functions but embarrassingly rudimentary in others. Boosters say this is part of a measurable progression towards AGI, which is not true, but is obviously helpful for marketing purposes. Because the apparent contradiction in proficiencies between something that charts whale speech accurately, and something that inexplicably writes “she made benches become men,” resolves fairly quickly when considering the different types of “cognition” involved.

As most “AI” skeptics are aware, Large Language Models function by identifying the statistically most likely next word / syllable / letter in a given syntactical structure (e.g., a sentence). They are trained on such massive data sets, and consult such massive data sets, that this averaging process can fairly reliably mimic the underlying logics. This is why they are exceptionally strong in pattern recognition, as in the case of the whales, because that is essentially LLM’s sole function: recognize a pattern, in order to mimic its progression.¹ Similarly, the ChatGPT advance on the geometry problem, “did not invent something fundamentally new that nobody saw coming,” according to one the mathematicians verifying its results, Sébastien Bubeck. “It just executed like an amazing mathematician.”

I will not pretend like I can offer anything approaching original commentary on the mathematics involved, since my actual understanding of math stops around Calc 2 and the rest is filled in from pop physics explainers and questionably executed biopics. But basically, GPT brute-forced a tactic that no one confined to a working memory and limited earthly hours would attempt. Contrary to Sam Altman / Elon Musk / Google creeps’ claims, we will never see “Artificial General Intelligence,” because the entire concept is a category error.

We know this is true because it can’t write a metaphor. As any middle school English teacher will tell you, metaphors function by connecting two seemingly unrelated objects in order to accentuate some other logic. Bear with me for a second as we explicate some gratingly simplistic examples: “he looked at the meat like a wolf.” That’s a simile, fine, but “he” and “wolf” are presented as different enough to need comparison, though share an underlying logic of hunger, and the comparison is steeped in unexpressed connotative associations typically (at least within certain cultures) ascribed to wolves: hunger, leading to desperation, and consequently ferocity, which leads away from empathic relation, and, if you’re a dime novel writer from a European tradition (or something), the comparison may signal a fundamental difference between man and animal (stupid) or civilization and barbarity (evil) and a reminder that “human nature” is corruptible, or something. All of that information is not only encoded “logically” within that very simplistic simile, but the construction also exists to relay a separate type of knowledge which undergirds the first: namely, feeling.

You, like all conscious beings, need to receive accurate information about the world to survive. It’s actually a pretty good way to be certain that your sensory information, even if incomplete, offers you a roughly accurate representation of external reality, because if it didn’t, you’d get crushed by trucks, constantly. If living creatures could not accurately perceive external forces like mass, velocity, shape, we wouldn’t survive long enough to reproduce. We don’t often enough consider “feeling” or “emotion” to be a comparable form of information.

Obviously, fear teaches us certain strategies: avoidance, aggression, something in between. Bring a rescue dog around unfamiliar men and you can see that the product of their basic pattern recognition assembles into feeling, the pituitary gland secretes its admixtures, and all of this primes towards certain behavior, and not just reflexive action, but subsequent perception.

Show me the amygdala on ChatGPT. Show me where, inside “DeepMind,” certain thresholds of understanding produce feelings of beauty, so I want to follow them further, ultimately leading me to discover new knowledge. My point is not, as the self-help-ish AI skeptics like to claim, that “love is what makes us, us <3”, or that there is an irreducibly human quality that magically ropes off our consciousness from machine replication. My point is that consciousness relies on physical structures and chemical processes that are entirely absent from the language calculators used to annihilate Palestinians indiscriminately and later claim they were terrorists, and that the only people it benefits to pretend otherwise are the natsec creeps who want you in the panopticon so they can make the Israeli definition of “civilized” a global phenomenon.

Look. Here is an English translation of Aimé Césaire’s Notebook of a return to the native land (emphasis mine, for later discussion):

Beat it, I said to him, you cop, you lousy pig, beat it,
I detest the flunkies of order and the cockchafers of hope.
Beat it, evil grigri, you bedbug of a petty monk. Then I turned
toward paradises lost for him and his kin, calmer than the face
of a woman telling lies, and there, rocked by the flux of a
never exhausted thought I nourished the wind, I unlaced the
monsters and heard rise, from the other side of disaster, a
river of turtledoves and savanna clover which I carry forever
in my depths height-deep as the twentieth floor of the most
arrogant houses and as a guard against the putrefying force
of crepuscular surroundings, surveyed night and day by a cursed
venereal sun.

At the end of daybreak burgeoning with frail coves, the hungry
Antilles, the Antilles pitted with smallpox, the Antilles dyn-
amited by alcohol, stranded in the mud of this bay, in the dust
of this town sinisterly stranded.

At the end of daybreak, the extreme, deceptive desolate eschar
on the wound of the waters; the martyrs who do not bear witness;
the flowers of blood that fade and scatter in the empty wind
like the screeches of babbling parrots; an aged life mendacious-
ly smiling, its lips opened by vacated agonies; an aged poverty
rotting under the sun, silently; an aged silence bursting with
tepid pustules, the awful futility of our raison d’être.

The reason that an LLM can pretend to be a cartoonishly racist version of tropical consciousness, but can’t make a metaphor make any fucking sense, is because it, is, making, averages. It can average a Césairean lexeme (or any brilliant writer it’s been trained on), but if you don’t understand the feeling of anger, of hatred even, of violence desired after violence wrought of the colonized subject, you cannot see how a house can be “arrogant,” you cannot see the psychic linkage between beauty and violence that forms the symbolic encapsulation of the colonizer, the “flowers of blood” which bring so-called civilization in quaint architecture all contrary to the pre-existing landscape made comparatively evil, the “cursed venereal sun” under which Antillean inhabitants have been “dynamited by alcohol” in a longstanding misery, “an aged poverty rotting” with no cultural function to express this consciousness in a form visible to its colonial master and so instead sitting in a “silence bursting with tepid pustules.” If you do not have the feeling-as-knowledge that constitutes these symbolic relations, they appear random, and thus all metaphor appears only as function of striking dissimilarity: as the LLM would have it, “the girl smiled like sunrise over a sink,” or “she had the kind of walking that made benches become men.”

Is it possible that one could create insanely complex connotative webs, massive word charts capturing all possible interrelations of language, and then train an LLM based on flagged language to intimate which poetic sense of “arrogant houses” is being used? No. The language is too tied to specific circumstances, so the problem becomes amusingly Heisenbergian; namely, any attempt to measure and contain such a connotative web would immediately find itself irrelevant upon completion, as language shifts and responds to history too instantaneously and subterraneanly to be meaningfully captured outside of poetics (expansively defined), but further that poetics have their own role in creating said web, meaning (bear with me) that each new instance of connotative creation this web attempts to measure will shift the measurement itself. Q.E.D.: no.

The limits to “artificial intelligence” are the exact same as its strengths––it can compare obscene numbers of whale language samples and spot correlatives, and it cannot make a meaningful metaphor, because it can never be conscious. Consciousness is product of material it does not have.

1. There’s probably an interesting exploration to be had in charting how this notion of intelligence is essentially identical to the shape-rotating IQ tests that Silicon Valley types obsessively praise (usually so they can be racist).