Follow the data, three years later

Three years ago I wrote an essay arguing that the old advice, follow the money, was out of date, and that you should follow the data instead. The thesis was that power had moved from owning assets to occupying the orchestration layer, that data wasn’t a commodity to sell but a form of foresight, and that prediction was just a slower word for control. I compared it to Asimov’s psychohistory and said, a little breathlessly, that we were building it as an advertising engine.

I want to grade that honestly, because the prediction machines I was gesturing at vaguely in 2023 have since shipped, and reading my own homework three years later is a useful and slightly humbling exercise. Some of it I called. Some of it I had exactly backwards. And the part I got most wrong is the part I now spend my days inside.

what held up

The core move held up better than I had any right to expect. Power really did keep migrating from the thing to the layer above the thing. The companies that mattered most over the last three years still own startlingly little in the physical sense. What they own is the position, the place in the middle where they can see the most and route the most, and the moat is still made of data, not assets.

And the prediction-is-control point aged into something more literal than I meant it. In 2023 I was talking about recommendation engines nudging you toward the next purchase. That was the toy version. The grown-up version is a system that can model you well enough to anticipate not just what you’ll buy but what you’ll say, what you’ll want, what you’ll do, and then quietly arrange the option you were already going to choose to be the one in front of you. I called that soft coercion. I undersold how soft. The best version of it doesn’t feel like coercion at all. It feels like the world being conveniently arranged in your favor.

what I had backwards

Here’s the part I got wrong, and it’s a big one. In 2023 I wrote about the predicting entity as a watcher, something that observes you, models you, and acts on the model. A surveillance frame. The thing that’s actually arrived is not a watcher. It’s a participant.

The models didn’t just learn to predict behavior. They learned to take actions, hold a goal, reach for a tool, do the next thing. That’s a different animal entirely. A surveillance system extracts. A participant operates. The 2023 fear was that someone would know me well enough to manipulate me. The 2026 reality is that I now hand a system goals and watch it go execute them on my behalf, and the unsettling part isn’t that it’s watching me, it’s how completely I’ve come to rely on it doing things I used to do myself.

I had the polarity reversed. I thought the danger was being seen. The actual frontier is being replaced, in the narrow, useful, almost invisible sense, one task at a time, by something I invited in and would not give back.

the part the 2023 essay couldn’t see

The thing my younger self genuinely could not have predicted is that I’d end up building in this layer, and that being inside it would make me less dystopian, not more.

From the outside, prediction-as-control reads as pure threat. From the inside, where I actually work now, the same capability is mostly just leverage, the boring, profound kind. A system that can hold a goal and reach for tools is, yes, the thing that could quietly run a population. It’s also the thing that can finally do the grunt work that’s kept whole categories of useful human effort locked up for decades. The same engine that could harvest you can also unblock you. The technology was never the villain. The thing pointing it has always been the only question that mattered, and I said that in 2023, and it’s the one line from the original I’d keep without changing a word.

the conversation that made it real

I want to tell you about a conversation I had this week, because it’s the moment the abstract thing in this essay walked into the room and sat down.

We were working through how to build a new product on top of an existing company, the unglamorous part, bridging what already exists with what comes next, cap table and all. And we got to the question every plan eventually hits: how many people does this take. How do we staff it. What’s the headcount.

And we both stopped, because the honest answer had quietly changed under us while we weren’t looking. We don’t need a big team anymore. We need one human with an army of agents per business unit. One person who holds the intent, the judgment, the accountability, and beneath them a layer of participants that can hold goals and reach for tools and just do the work that used to require ten hires and a quarter. That’s not a forecast. That’s the staffing plan. We said it out loud and then sat in the small silence that follows saying a true thing.

Here’s why it chilled me, and I mean that precisely, not as a figure of speech. This essay is an abstract argument about prediction becoming a participant. That conversation was the argument made flesh, in my own work, about my own headcount, on a Tuesday. I wrote in 2023 that the danger wasn’t being watched, it was being replaced, one task at a time, by something invited in. I did not expect to be sitting on the replacing side of that sentence, doing the inviting, drawing the org chart where most of the boxes are no longer people. The thing I diagnosed from a safe distance is now the most efficient way for me to build, and I’m reaching for it, because it works, and the part of me that wrote the warning is watching the part of me that’s executing on it, and they don’t fully agree.

The synchronicity is almost too clean. The kid who wrote that prediction-is-control would be a little afraid of the builder who just casually designed a company around one-human-per-unit and called it efficient. He’d be right to be. That’s the cost of being early to your own future: you get to watch yourself become the thing you saw coming, with your eyes open, choosing it.

So: partial credit. I saw the layer move. I saw prediction become power. I missed that prediction would learn to act, and I missed that I’d be one of the people teaching it how. The watcher became a participant, and then the participant became a colleague, and now I’m not sure where I end and the orchestration layer begins, which is either the most interesting place to be standing or the first symptom of the exact thing I warned about. Probably both. Follow the data. It went somewhere even I didn’t expect, and I was already looking.