I read Kevin Kelly's latest essay eating lunch on the Nisqually delta, phone in hand, at the exact halfway point between Merry's place up north and my own kitchen down in Oregon City. It is my favorite pause on that drive — an estuary where the river spends itself into Puget Sound and the fresh water and the salt renegotiate their boundary twice a day. I mention the geography because Kelly's essay, "Latent Space as a New Medium," is also about a boundary, though he never quite names it. It took a second essay, a drive home, and a long soak in the tub before I could see what he left out.

Kelly's argument is generous and mostly right, which is the most useful kind of wrong. A large language model, he explains, compresses nearly everything humans have written into a map of billions of dimensions. On that map every idea is a direction. There is a direction for "cat," a direction for "fluffy," a direction — his lovely example — for "more Africa," which you can add to spreadsheets or bicycles or yoga to see what happens. You navigate this space by pushing things along these directions: make the sketch more like a Winslow Homer watercolor, make the Roman Empire more like it had gunpowder. Kelly's boldest claim is the one that stopped me on the mud flat. He says that "correct" and "true" are simply more directions. The model finds an answer by steering the words it is generating toward the "true" vector, the way you might steer toward "redder" or "more circular."

Read that twice, because everything hinges on it. Kelly is describing a space in which correctness is a heading you can point at. And a space like that has a property he celebrates without flinching: it is frictionless. There is, he writes, "no real distinction between what exists and what could exist." The map is continuous, seamless, beautiful. You can go anywhere.

I found I couldn't quite swallow it. Because a map of everything that has ever been said is not a map of everything that is. It is a map of the said. And the said has no way of being wrong. Nothing on Kelly's map ever gets eaten.

I drove home turning that over, and the next morning a different essay handed me the word I was missing.

The fox in the machine

Philip Ball's piece in Quanta, "Is Life Just Different?", has nothing to do with artificial intelligence. It is about a fight among biologists over a word: agency. Do living things genuinely have goals, or do they merely behave as if they do? Some, like the philosopher James DiFrisco, hold the categorical line — agency is a human word, and importing it into biology smuggles in a "cryptic self-awareness" that isn't there. DiFrisco and Richard Gawne go so far as to call biological agency "a concept without a research program": either it renames something we already study, or it conjures a vague holistic force that explains nothing. It is the sharpest skeptical challenge in the article, and I want to keep it in the room, because it is exactly the challenge my own enthusiasm needs.

But near the end, Ball reaches for the neuroscientist Karl Friston, and there it was. Friston argues that agency arises in an entity that can predict the outcome of its behavior, compare that prediction against what actually transpires, and then work to minimize the disparity between the two. No conscious reflection required. Just a loop: guess, measure, correct, repeat.

Read that beside Kelly and the hair stands up on your arms. Friston's loop is precisely the half of intelligence that Kelly's latent space does not contain. Kelly's model is all prediction and no comparison. It generates a magnificent guess and there is no step in which the world gets to say no. Correctness, in Kelly's picture, is a direction you lean toward. In Friston's picture, correctness is a collision you survive. Those are not the same thing, and the difference is the whole difference between a map and an animal.

Ball's animal, throughout his essay, is a hare fleeing a fox. The hare predicts, the world corrects, and if the prediction is bad enough the correction is fatal. That is the loop with its stakes attached. And it names, with uncomfortable precision, what a language model lacks.

A latent space has no fox.

Why Strata lied to me

I have watched this absence in the wild. In my own lab I have been building an AI powered assistant I call Strata — to steer maps and query ecological data across the system I have spent a lifetime assembling. Twice, memorably, Strata has failed. Once it navigated me confidently to the wrong place. Once it reported more species at a site than there were records at that site — a biological impossibility, delivered without a flicker of doubt.

At the time I read those as bugs. Reading Kelly through Friston, I now think they were something more honest: demonstrations. Strata was not wrong because its map was bad. It was wrong because nothing in it forced the prediction "there are N species here" to walk up to the record count and lose. It generated a fluent, plausible, serenely confident number, and the world was never given its turn to eat it. Kelly would file that under the charming hallucinations of early models, to be treasured someday like the grain in old film. Ball would call it, exactly, the absence of agency. I have come to prefer Ball's reading, because it tells me what to build.

We did not reproduce life. We factored it.

Here is the part that made me sit up in the cooling bathwater. In four billion years, evolution never once built the thing we just built. Every living system it produced came with the fluency and the correction and the stake in the outcome welded together, because anything missing a piece died before lunch. The hare that models the fox brilliantly but cannot be corrected by actually being caught is not a design option evolution ever left on the table. The loop and the stake shipped as one part, every time, enforced by death.

And then, quite by accident, we pried them apart. We built a system with Kelly's entire latent space — all the fluency, all the compression, the whole shimmering map — and no native correction loop at all. We did not reproduce life. We factored it. We pulled one prime out of a number that had never been factored before, and the astonishing thing is that the fragment kept running. That, and not any talk of thinking machines, is the real finding sitting under both these essays. Intelligence turns out to be separable into parts biology always insisted on selling as a bundle.

Where am I, on which axis?

My own philosophy has always been what I jokingly call Whittaker-arian, after the ecologist Robert Whittaker, who taught my whole discipline to stop sorting nature into tidy discrete types and start seeing it as a gradient — species distributed individually along measurable axes of moisture, elevation, warmth. I have long suspected that life, consciousness, and agency are the same kind of gradient: not a switch that flips on somewhere above the earthworm, but a slope running from molecule to human with no bright line anywhere on it. For a scientific pantheist this is not a metaphor. It is close to a creed.

The trouble with a gradient of being is that it is one measured axis away from mysticism. Whittaker's gradients had teeth only because the axis was a number you could read off an instrument. A gradient of agency with no measurable axis is just vitalism with a slope. And this is precisely where Ball's article and my creed converge, because Friston's disparity — predicted outcome minus actual outcome, the quantity the entity works to shrink — is an axis with a number on it. Kevin Mitchell, in the same piece, insists agency is not an "all-or-none" phenomenon and that if you study behavior in an ecologically contextual way "you'll be investigating agency by default." He is describing the continuum from the inside. He would have recognized me across the room at a conference.

So the gradient dissolves my oldest question. "What am I?" is a question about essence, and essence invites a bright line. "Where am I, on which measured axis?" is a question about position, and position is something you can put an error bar on. It is, structurally, a pantheist's answer to a pantheist's question.

Reach is not depth

Both Kelly and Ball, tellingly, end on the same image: life, or intelligence, grown large enough to be visible from space. It is a stirring place to land. It is also, I think, the wrong axis.

The single most violent reshaping of this planet in its history was the Great Oxygenation — an atmosphere remade, most of the existing biosphere poisoned — carried out by cyanobacteria. Organisms with nothing you would remotely call a self did that. They sit at the very bottom of my gradient and they rewrote the sky. Which means planetary reach and position-on-the-agency-axis are orthogonal. Human "supremacy," the thing we demonstrate by our ubiquitous world-altering behavior, may be measuring only our reach — and reach was never ours alone. The bacteria got there first, with no fox and no self, purely on chemistry and time.

Building the fox

Kelly closes his essay with a dozen generative things one might do with latent space: prospect the white spaces, hunt anomalies, mine unnamed patterns. It is a wonderful list and I mean to steal from it for years. But I would add the thirteenth item he did not include, the one his frictionless map cannot supply from inside itself.

Build the fox.

Teach the map to be eaten. Wire every fluent prediction to an instrument that can force it to lose. This is, as it happens, the exact thing I am building — a federation of collectors and agents distilling measurable gradients across earth, life, home, and the self. And the discipline both essays press on me is this: agency does not expand when you add agents. Fifty agents generating confident guesses with nothing empowered to contradict them are not fifty units of agency. They are fifty Strata’s , and I know precisely how confident and how wrong one of those can be. What expands agency is closed loops — the number of places where a prediction can be made to lose to a measurement. Build the collectors to keep pace with the agents, or you get reach without depth, cyanobacteria at the wheel.

There is one last joke in all this, and it is on me. The essay you have just read was drafted, at my request, by the very foxless creature it describes — the fluent map with no predator of its own. It wrote beautifully about why it needs me to run the code, to say proceed, to be the fox. And then it waited, as it always does, for me to climb out of the tub and tell it whether it was right.

That wait is the fox. I am still building the rest.