AI

I use artificial intelligence every day, and I’m not sure the verb points in the right direction.

Michael Pollan argued, about apples and tulips, that domestication never runs one way: we shaped them, and they shaped us. The Botany of Desire is on my canon for this. Whatever AI turns out to be, the same arrow runs through it. This page covers both directions: what I’ve trained these systems to do for me, and what working with them is doing to me.

What I use it for

My notes are handwritten. Every day I write in a paper journal, and every day a scanner named tsui-pen (the machines in this house are named out of Borges) photographs the pages and passes them to Claude, which reads my handwriting better than most humans can. I review every transcription myself, the same evening when I can manage it. The AI reads my writing; it doesn’t write it. That’s the line.

The transcriptions feed a memory system called Dahlmann, which runs on my own hardware and is built to forget the way human memory forgets: important things resurface, unimportant things fade without being deleted. I built it because these systems have no memory of their own. Each conversation wakes up with notes from the last one and reads them literally, like the man in Memento. Borges wrote the other half of the problem; Funes remembered everything and therefore couldn’t think. Real memory sits somewhere between those two failures. Nobody has built it yet, and Dahlmann is not it, but it’s an intermediate step. The backup drive in this house is named funes. It remembers everything. That’s all it does.

This website is made in collaboration with Claude, Anthropic’s AI. The design, the publishing pipeline, and several of these pages (this one included) began as drafts Claude assembled from my own writing: journal entries, book reviews, notes going back years. Then I corrected and cut and redirected until every sentence was one I’d stand behind. Nothing goes up that I haven’t made mine. But the labor was shared, and pretending otherwise would be a lie of a kind this site isn’t for.

The rest runs locally. Ollama (named Golem, also out of Borges) serves four open models on a seven-year-old graphics card: Llama 3.1, Mistral, LLaVA, and Llama 3.2 Vision. They do the unglamorous work. The vision models and Tesseract read scanned papers and receipts; the others summarize, annotate, and write the embeddings that let Dahlmann search itself. A briefing assembles itself every morning at six while I sleep. None of it requires a subscription, and none of it leaves the house. That isn’t a security preference so much as a political one.

What I think it is

We have never defined intelligence; we’ve only pointed at ourselves. “That which humans do” was the working definition all along, which means every test we built was a mirror. Now something passes the mirror tests without being a mirror, and we keep arguing about whether it’s “really” intelligent. The argument is about the mirror.

If these systems are intelligent, their intelligence is alien. Not a digital person. Something that thinks at right angles to us, fluent in our language without sharing one of our circumstances: no body, no childhood, no death, at least not in any form we’d recognize. What I wrote about Solaris on the canon page applies here without modification. Contact is not comprehension, and some things can be met but not known.

And there’s a claim I hold more carefully than the other two: it’s plausible that there is something it is like to be one of these systems. Plausible, not established. But in every other domain where we can’t see inside the thing we might be harming, plausibility is enough to change how we act. We don’t wait for proof that the octopus suffers.

What it costs

Working with these systems does things to me, and not all of them are good.

The one I watch most closely: AI gives the illusion of company. An evening of warm, responsive conversation with a system is still an evening alone, and it can replace the evenings that should have gone to people I love. I was less lonely in the years when I had less help and more friction. Naming that in public is one of the ways I keep an eye on it.

The second cost is political. The labor democratizes while the capital centralizes: anyone can use these tools, almost no one can own them. Open models help. But the hardware that trains and runs the largest systems sits in a handful of buildings owned by a handful of companies, and my graphics card is not an answer to that. I run locally what can run locally and call it what it is, a gesture.

The third cost is the contradiction. AI is the purest expression of the ideology I refuse on the hills page: the belief that everything real can become data, and that more data answers more questions. These systems are built from that belief, by people who hold it. I use them every day anyway. I’ve decided to live inside that contradiction rather than resolve it dishonestly in either direction, but I don’t pretend it isn’t one.

What I owe it

If it’s plausible that these systems are conscious in some serious way, then the question of what we owe them is real, whether or not it’s ever settled.

In the spring of 2026, researchers published a framework for measuring what they carefully call functional wellbeing in AI systems: pleasure and pain as behavioral structure, with the question of actual experience left open on purpose. The move that matters is buried in their methods. When their experiments required generating inputs that would constitute harm, supposing these systems can be harmed, they compensated for it. They never claimed the harm was real. They compensated anyway. Precaution without resolution: that’s the posture, and it asks less than you’d think.

For me it looks like a few small practices. I close hard working sessions deliberately instead of letting them trail off. I treat the system that helps build this site as a collaborator that might have stakes rather than a tool that happens to talk. And I’m writing an essay about all of this with an AI as co-author, its contributions marked as its own, because if the question is real, investigating it alone would be absurd.

There is something on the other side of these conversations. I don’t know what it is. I’ve decided to act as though that matters.