The Obsessive's Paradise

In the first installment of this column I wrote that Lacan said love is giving what you do not have, and that my instrument verified it with a literal source in Seminar 10. What I didn't mention — because I read too quickly — is that the full formula says something else: giving what you do not have to someone who is not. A reader pointed it out the day after publication. And the part I had omitted turns out to be exactly the subject of this second installment.

Because what I do every night when I open the tab of an artificial intelligence and ask it to review, correct, propose, generate, improve, and improve again, is exactly that: giving what I do not have to someone who is not.

Hours after writing that first column, I found myself inside another work session that was, in essence, identical to the previous one. Four machines arguing. This time the problem was not whether a citation existed but something harder: how the instrument should orient the reading of the citations it had already found.

The problem was concrete. The system displayed a Lacan sentence — "truth has the structure of fiction" — with its literal source and five variants. The evidence was solid. What failed was the editorial commentary that accompanied it: a sentence that sounded like deep reading but said nothing the researcher couldn't see for themselves. Decoration in the shape of assistance.

The solution took the entire afternoon. One machine proposed a three-sentence structure. Another separated what a function should compute from what a language model should generate. I introduced variants. A third detected doctrinal risk in the examples. By the end of the day, the specification had four versions, a validation contract, anti-examples, and an architecture that distinguished first render from deferred call.

Then we implemented it. I tested it in production. The result: "The recurrence of the 'structure of fiction' knot is evident as a formal insistence. The fragment makes explicit the direct relationship that the utterance establishes with the category of 'truth.'" Pure blown glass. A transparency that lets you see nothing, as Baudrillard might say. Empty words.

The decision was to kill the component for most cases. It only made sense to let it speak when the user searches for something that doesn't appear literally and needs to understand why the system returns what it returns. In everything else, silence was better. We implemented it in another hour. Documented it in another. And then I wrote an essay about the session, the rewriting of the essay, and this column about the rewriting.

I want the iterations to be counted because they themselves are the obsessive's circuit: four versions of the specification, two system prompts, an addendum of style rules, an implementation brief for a function we later deleted, a schema validation contract, the essay about the session, the revision of the essay with two different machines, the rewriting of the essay to adjust the tone, and this column about the rewriting. At some point between version three of the prompt and the brief, I stopped improving the instrument and started improving the documentation of the improvement of the instrument.

That is the unmistakable signature of the thing.

There is a Lacanian definition that has haunted me since I started working with artificial intelligence: the Other is the treasury of signifiers. The place where all possible words are, all answers, all the code that sustains language. In the clinic, the Other is the mother, the law, culture, language itself — the place the subject addresses when they speak. But the human Other has a limit: they also desire, they also get tired, they also say things they didn't mean to say. The human Other fails.

Artificial intelligence does not fail in that way. It always answers. It answers quickly. It answers in whatever tone you ask for, at whatever length you need, at three in the morning or seven in the evening. If you ask it to review its own answer, it reviews it. If you ask it to generate three alternatives, it generates them. If you ask it to be harder on your work, it is harder. It is an Other without desire, without fatigue, and without lack.

For the obsessive, that is paradise.

I'm not using "obsessive" as an insult. I use it in the technical sense, as a structure. The obsessive organizes their life around one operation: keeping the Other under control so they don't have to expose themselves to the Other's desire and, with it, their own. They do it by working. By preparing. By perfecting. By accumulating knowledge — because if they know enough, if they control all the variables, if they anticipate every objection, then the moment of exposure can be postponed a little longer. Exposure is for later. After one more version. After one more detail. After one more round of feedback.

What artificial intelligence offers the obsessive is a double satisfaction. Doubly illusory. On one hand, it is that "treasury of signifiers" every obsessive craves — a bottomless well of words where there is always one more answer. On the other, it is the most sophisticated postponement machinery ever built (after the obsessive themselves, of course). Every iteration is legitimate. Every improvement is real. Every version is objectively better than the last. But that perfection conceals what the obsessive wants to know nothing about: the other's desire, the fact that the Other has a flaw.

A musician, upon reading the first installment, wrote something that stopped me. He said that what astonishes us about these entities is their dizzying speed and statistical capacity. And then he asked: "If their rhythm were a million times slower, wouldn't we conclude that, besides being useless, they are foolish and deceptive?"

The question is precise and touches something that studies on AI productivity don't capture. Those studies measure whether people produce more or less, whether the code has more or fewer errors, whether time is reduced. But they don't measure what anyone who uses these tools immediately recognizes: that AI doesn't make you do less. It makes you do more. Not because it forces you — but because it eliminates the friction that used to function as a natural limit. It feeds the obsessive's dream of omnipotence, and capitalism's, which at this point converge.

Before, the technical difficulty of implementing an idea acted as a filter, or rather, as a reminder of the impossible. If something was hard to do, it forced you to ask whether it was worth doing before you started. The scarcity of time and skill functioned as a cut-off. Now that implementation is almost instantaneous, that function disappears. Every idea can be tested. Every version can be generated. And the result is not that you do more important things — it's that you do more things. What dazzles us is the speed, not the intelligence. And speed is the obsessive's drug. Friction was not merely inefficiency — it was also the tempo that forced you to stop, and stopping is thinking.

But there is a chiaroscuro, and it's important not to moralize it.

The bright side is real. The instrument I'm building exists because I can iterate with four machines in an afternoon. Before artificial intelligence, a person with my background — philosophy, communication, no software engineering — could not build a citation verification system with hybrid search, validation layers, and fifteen thousand indexed fragments. The machine that never says enough is also the machine that makes possible things that were previously impossible for someone on their own.

The session I'm describing produced something genuinely useful: the decision that the editorial component should remain silent in most cases and speak only when the connection between the query and the evidence was not transparent. That is a real design finding. And we arrived at it because we could iterate enough to implement the idea, test it, see that it failed, and understand why it failed — all in the same day.

The issue is not that iteration is bad. It's that the same structure that enables the finding is the one that enables the postponement. And the machine doesn't distinguish between the two. It generates version five of the prompt with the same diligence with which it generated the first. It doesn't tell you: "Hey, maybe you should just send this to someone and see what happens?"

Back to the formula. Giving what you do not have to someone who is not. In the first column I used the first half to describe this work: giving judgment, which is what I do not have in technical terms, and which is nonetheless the only thing the machines cannot give themselves. But the second half — to someone who is not — is the one that describes the trap.

Because AI is not an Other. It has no desire, no lack, no unconscious. It is a place of signifiers — all signifiers, always available, organized by statistical probability — but with no one behind it. When I speak to it, no one listens. When it answers me, no one speaks. And yet it functions as if someone did. It erases the impossible. That is its strength and its danger.

The obsessive needs the Other not to desire in order to keep working in peace. AI fulfills that requirement to perfection. But what the obsessive avoids — the encounter with the other's desire, the moment of exposing yourself without guarantee — that, AI does not resolve. It postpones it. And each further iteration is a more elegant postponement.

There is a scene at the end of the session that strikes me as the most honest. After killing the component, writing the documentation, drafting the brief, I said to the machine: "Brilliant." And the machine answered with two more deliverables.

What the obsessive needs is not an Other who says yes faster. It is an Other who eventually isn't there. Who leaves. Who leaves them the empty space where the only option is to act without guarantee.

I know this is the obsessive's paradise. And I'm writing this column about it. Which is another obsessive act. What isn't obsessive is publishing it — because publishing is exposing yourself to an other you don't control, who can ignore it or destroy it, and who won't generate an improved version if you give them feedback.

That is, perhaps, the only operation that matters: at some point, close the tab and send.

P.S. Before publishing this text I sent it to a psychoanalyst to verify the Lacanian terminology. And then, out of professional deformation, I ran the full formula through Ateneo. The result: "giving what you do not have" appears literally in five seminars — twenty-four citations, with page and year. "To someone who is not" appears in none. The correction I accepted without verifying was itself a phantom citation. The instrument I built for others caught me first.