asemic horizon

I.

The world’s shortest critique of Pavlovian reason: if the pointer address of dopamine response can slide backward from stimulus (beef) to cue (bell), then it makes sense to start looking at every particular experience as cue to a parallel, abstract, universal real counterpart. Emancipation is seeing the cues for themselves, their this-ness, removed from the great chains of salivation.

II.

By now, augmented LLMs (verbal reasoning, post-training witchcraft, agentic loops) have rightfully inherited the mantle of “artificial intelligence”. Large Language Models in their own right were originally parlor tricks; they produced strings of text on which we, the users, hallucinated meaning. The real bitter lesson that ensued was that the titanic scale of corpora offered a cheat out of Hubert Dreyfus’s Heideggerian critique of AI — effectively, LLMs gained a world, albeit one that’s only truly alive during post-training.

What is alive to the user is, of course, the world of tools to be called; this allowed “frontier AI” to subsume all extant “artificial intelligence” paradigms (from logic programming and constraint satisfaction to task-defined machine learning). Everything available that’s funneled into “context engineering” (the original parlor trick) sort of fakes a kind of worldly learning. But the limits to post-training (which are more or less coextensive with the limits to “artificial intelligence” in previous paradigms) are the whole beauty of it — past an ambient level of functionality which is invisible to us (“what is water?” asks the fish), everything that’s locked-in during post-training is a source of discomfort. This, of course, is the drama of policy more broadly: la carapace de la tortue, outile et pénible; we live in a society, the rules of which both oppress and liberate.

III.

The architecture of “frontier AI” inverts many assumptions of postcognitivist vulgata. Broad cognition, which the old heuristics had us to locate in the broad external networks of affects affecting affects, instead resides inside the faux-brain of LLMs; specific, left-sided, analytical reasoning is in the technological periphery. The salient example: the real site of mathematical skill behind the recent spat of bravura display of open-problem solving by frontier AI is in traditional theorem provers; the race to expand the mathematical reach of machines is coextensive with the race to complete said theorem provers. Frustratingly, frontier AI is “the idea guy”; we (the user, mostly but also the growing necklace of tooling around LLMs) tasked with providing precision.

The light at the end of this dis-humanistic tunnel? Well, we can use AI to help us produce said tooling. This is the fast “recursive self improvement” loop; it grows in hourly leaps even as the labs play up the drama of AI taking over the development of the language model. In the future (and the whole anxiety script around the collapse of futurity is that the present is a thing of the past; that rational people live in the margin, that we can no longer live anywhere but in the future), we’re all paperclip maximizers for 15 minutes or so.

IV.

What computers can’t do: concept violence. Take the idea of consent in sexual morality; it can best be qualified as brutally idiotic — idiotic because it looks at human engagement as textual rather than textural, and brutal because it attaches itself to the heinous crime of rape, sin against nature. Search your sexual memories, try to pinpoint the threshold of consent — it isn’t there! At best, consent is a fetish, a delicious ingredient of dominance and submission: unlike the more “naturalistic” kind of sex where events feverishly unfold from events, submission-fetishistic sex explicitly pauses at a threshold where — in jest — ordinary dignity is suspended. It goes without saying (as a consensus belief in modern sexual culture) that the consent that enables play-debasement has nothing to do with rape. Of course, you can’t consent to rape; whatever you explicitly consent to is already a kink.

V.

From its very ground, “democracy” as defined by the vote is descriptive, not prescriptive. It says that what most people want should carry the day, like scientific linguistics says the rules of language are determined in usage. But then, why vote for someone in particular? Worse yet, why participate in political discourse, if not to somehow influence outcomes? Taking a political side means taking sides on what people should want.

This is how politics is the continuation of war by other means, and not the other way around: a more civilized form of conflict, it is conflict regardless. To say that democracy is an end unto itself is like saying that the particular size of football pitches is a fundamental truth of the game — whereas the truths of sports are discovered in play, in the this-ness of movement. Let us no longer speak of democracy or consent.