← Data Into Labor · All essays

Data Into Labor · A Note Before the Essays · V2

The Thoughts Are Mine. The Sentences Are Ours.

July 2026 · AJ Orbach

Every essay on this site was written with AI. I want to say that plainly, before I ask you to read any of them.

When people hear that, most of them quietly downgrade the work. The reasoning goes: sentences are how thinking shows up on the page, so if a machine made the sentences, the machine must have done the thinking, and the name at the bottom is decoration. For most of history that reasoning was sound. The writing and the thinking were done by the same person, because there was no other way to get writing.^[1]

What I want to explain in this note is why that reasoning fails now, at least for these essays. It fails because the two things it fuses, having thoughts and making sentences, were never actually the same skill. I happen to be an extreme case of the difference.

Voice

I'm dyslexic, and I'm bad at writing. I don't say that to be modest. I mean it literally: I never learned to make prose do what I wanted.

When I write on my own, I write from inside my own head. I skip the bridge between one idea and the next, because from where I sit the bridge is already visible. I assume the reader has sat in the meetings, seen the numbers, watched the channel fatigue, and can feel where the argument is going before I explain it. People inside the business can mostly follow me. Everyone else experiences my writing as arriving in the middle of a conversation.

So for most of my life there's been a painful distance between what I could sense and what I could express. I could see a pattern across a hundred merchant conversations. I could hold an entire structure in my mind. And when I tried to put it on the page, the thought came out smaller, flatter, and less precise than it went in.

These models are the first tools that have made me feel I can actually express myself. I speak the thought in the order it arrives. I hand over the context I would normally skip. Then I steer: "No, that's too clean." "You lost the tension." "Go back to the merchant's P&L." "This sounds like a machine." I keep steering until the language holds the shape of what I meant. For the first time, the distance between thought and expression feels crossable.

I won't pretend the machine was a spell-checker. It did real work. It found language, built transitions, proposed structures, and sometimes showed me a connection I hadn't made explicit. The sentences are genuinely a collaboration. But finding language for a thought is not the same as having it, which raises the obvious question: where did the thoughts come from?

Weights

Not from reading. The essays here don't come out of decades of books. They come out of years of building, most recently Triple Whale: the data platform, the measurement stack, and now Moby, the intelligence that sits on top of both. Behind these essays are thousands of conversations with merchants, launches that worked and launches that failed, pricing arguments, attribution arguments, hiring, firing, near-death quarters, and quarters that felt like flight.

Very little of that survives as stories. You don't walk away from years of operating with a catalog of lessons you can recite. What the years leave behind is stranger: they change your instincts. They change which questions feel serious, which answers feel too easy, and which tensions you refuse to resolve cheaply. Whether a channel is actually incremental. What a merchant really means when they say they trust a number. What happens to a team when the work it sells becomes cheap.

I sometimes think about this in the language of the models themselves. My weights have been trained. Every merchant call, every failed launch, every dashboard argument, every board meeting adjusted something. What's left isn't a database of experiences. It's an intuition for the shape of a business and the moral weight of a decision.

That intuition is the raw material I bring to the machine. The machine helps me unfold it. It can't explain where it came from.

Loop

I should say what my seat is, since it bears on how much to trust me. I'm a cofounder of Triple Whale, and our work keeps us at the frontier of applied AI. We benchmark the leading models, and this generation, Fable 5 and GPT 5.6, is the first one smart enough to do the work we'd been preparing for.^[2] We built Moby, the harness that turns those models into a working teammate for e-commerce operators, and I use it as customer zero every day. These essays are written inside the machine they describe.

Living that close to the systems has taught me that the model is only one part of the result. The harness matters. The context matters. The sequence of prompts matters, and so does the source material, and so does the ability to keep a long argument alive across many rounds. Above all you have to know when the model has produced something fluent but empty, elegant but false, or confident but unsupported. The failure mode of these systems isn't gibberish. It's beautiful language wrapped around a weak claim.

So there's no magic sentence that produces an essay. The process is a loop, closer to a workshop than a prompt. I start with a question I actually care about, usually a tension I've carried through years of operating. I dictate the interior argument in fragments. I load the context: prior essays, strategy notes, product truth, tone, audience. On this site the model isn't guessing about the company; it's connected to it, and it can check an essay against the real system.^[3] Then I argue with the draft. I ask for counterarguments. I restore ambiguity where the prose has become falsely certain. I cut language that sounds impressive and says nothing. Sometimes I dictate most of the idea; sometimes the model finds the real structure; sometimes a draft shows me my argument is incomplete; sometimes the model is simply wrong. The useful work happens in the movement between those states.

And before anything goes out, I read a complete edition and decide it belongs in the series. Then I put my name on it.

Mine

It's tempting to look at all this and conclude that the work of writing has disappeared. Some of it has. The mechanical struggle to turn a dyslexic stream of thought into clean sentences is no longer the barrier it was for forty years. But the sentence is only the visible end of a long process. The company took years to build. The questions came from operating it. The intuitions formed slowly, across thousands of merchants and more mistakes than I'd like to count. The speed of production conceals the slowness of formation. AI compresses the last mile. It doesn't replace the road.

I should also say what these essays are not. They're not press releases and they're not a roadmap; nothing here is a promise of features, timelines, or results. The models can make mistakes, invent certainty, and flatten disagreements, and I will not catch every case. Keep your judgment awake while you read. I try to do the same while I write.^[4]

One more thing about the word "mine." When I say the thoughts are mine, I don't mean they appeared from nowhere. They're indebted to my cofounders, the team, the merchants who trusted us with their businesses, investors, and the competitors who kept us honest. "Mine" means I chose them, wrestled with them, and accept responsibility for publishing them.

The tools are new. The questions about work, trust, and value are old. And that, I think, is where authorship now lives. It used to live in the sentences, because the sentences were the hard part, and the only person who could make them was the person who had the thought. Now the sentences are cheap and the rest is still expensive: the years, the questions, the judgment about what's true, the willingness to sign. The thoughts are mine. The sentences are ours. The name at the bottom is there on purpose.