A lot of the sessions I sat in at beCamp this month were wary of AI assistants. It was Charlottesville’s tech unconference, celebrating its twentieth anniversary at UVA’s School of Data Science, and there was plenty of AI in use in the hallways. Even the couple of people I talked with who had largely opted out of a life tethered to a screen said they open Gemini or ChatGPT now and then.
So it’s everywhere, even among the people who’d rather it wasn’t. I don’t think that makes anyone a hypocrite. I think it makes this an old argument.
When Muddy Waters got to Chicago in the 1940s, nobody in a crowded club could hear an acoustic guitar, so he plugged in. The amplifier made him loud enough to be heard over the room. It also made every flubbed note audible, and he had to change how he played. When he toured England in 1958, many blues fans who’d come for an acoustic set were startled by how loud he was. A few years later a young London band named itself after one of his songs.
New tools always take a while to find their use. I’m 110% sure whoever invented the pneumatic framing nailer built several acres of unnecessary 2x4 assemblages before somebody used it to frame a house three times faster.
Nobody I know ever complained about spellcheck, or the grammar checker that underlines a run-on. The assistants are different. They make real improvements, from the choice of a single word up to the shape of a whole piece, and they leave fingerprints. You know the ones: the long dash in every other sentence, the word that’s a half-step off from the one you’d have reached for. Like the amplifier, the tool makes you louder and makes the mistakes louder too.
The suspicion of “slop” is fair. A board full of nails and screws that’s been on the wrong end of every power tool in the shop is worth less than it was as a tree.
I’ve written some things over the years that I was proud of, and a few that did real work: an essay or a cover letter that carried me past other applicants for a job at a Y Combinator startup or a place in a program in college. Every one of them took days. The days were the point. They put distance between me and the draft, so I could come back to it as an editor, or as a reader, and collaborate with myself. Even so, I skim my own phrases. Months or years later I spot the blemish, usually obvious and rarely worse than mildly embarrassing, and my eye catches on it forever after like a thorn in my sock.
For most of us, me included, an assistant is the editor who’s available right now, and it gets a draft to minimally competent without the days. It doesn’t get tired on the fourth pass. It doesn’t lack patience; it lacks the concept of it.
In 2017, eight researchers, most of them at Google, published Attention Is All You Need, the paper that gave us the transformer. Nine years on, I’m still blown away by what LLMs and the agents built on them can do, but what I’m realizing is that the superpower is the language. An LLM lets a computer read and write in our native language, in a way it couldn’t before.
Language is the one channel that carries an idea between any two minds, human or machine. And ideas compose. Very little that’s new is new all the way down; mostly it’s an untried combination of things that already existed. The more ideas get said, and said well, the more combinations there are to try.
Muddy’s early record “I Feel Like Going Home” lent its title to Martin Scorsese’s film in The Blues, a seven-film PBS series from 2003 about the music and the people who made it. Outside of efforts like that one, very little of their history was written down. The music outlived the record anyway. You can hear it in a mountain of what came next.
Ghostwriters have helped people tell their stories for a long time, but they write what pays: either the author can afford one, or a publisher believes the story will sell. There’s nothing wrong with that. It just selects for a small group of people and topics.
Every person has something to say that the rest of us don’t, some bit of insight that came from a life nobody else has lived. Our differences guarantee it. What’s been scarce is the means to say it well, and that just got cheaper.
What you write for your agent is what you know, written down. If you work with agents, you’ve probably written a fair amount of it this year without calling it writing: a skill that teaches an agent how your team ships a release, a CLAUDE.md that holds the house rules for a codebase, a workflow that took three tries to get right. Those are prose. A colleague can read them as easily as a model can, so they can be handed to another person or another agent, and combined with someone else’s into something neither of you would have written alone.
Right now most of it lives in a dotfile on one laptop, copied by hand when someone asks. That seems like a waste of good writing.