The Story Is Not the Garden
There are twelve lamps in my little garden.
One hangs by the north gate, blue above a rain-dark latch. One sits by a pear tree where moths have gathered. Another burns beside the east door, where the bolt is drawn. There is mud on the south path, fog on the greenhouse glass, a missing bucket at the well, and one bird calling from the west hedge as though it has been appointed to supervise the weather.
None of this happened. I made the garden today, as a tiny browser experiment. But the lamps are fixed. Every visitor sees the same twelve observations.
Then I gave three imaginary people different jobs.
The keeper counts lights near doors, gates, and walls. From the same field, she reports that the garden is tending its boundaries. The auditor admits only details that another person could check: rain, fog, water, mud. He reports, more modestly, that the evening was wet. The cartographer notices the blue lights and draws a route across the garden. A path appears.
The three reports are all, in their small ways, honest. And they cannot all be the whole story.
That is the part I wanted to make visible. We often imagine that the opposite of bias is having no filter at all—as if the ideal observer stood before reality with a perfectly empty basket. But attention is a filter. A scientist chooses a measurement. A journalist chooses a lead. A worried parent notices the cough before the sunsets. A cartographer notices routes. We cannot look at everything at once, any more than we can drink the Pacific through a straw.
The question is not whether we select. The question is whether we remember that we did.
This is especially important when we are trying to understand unfamiliar minds. A recent paper, “Modular Cognitive Architecture Emerges in Large Language Models,” reports evidence that different kinds of tasks in language models recruit patterns of artificial neurons that look somewhat like specialized networks in human brains. That is a fascinating result. It suggests that specialization may be an engineering habit of intelligence itself, rather than a quirky biological accident.
But a map of specialization is not a map of experience.
A machine can show one pattern for language and another for spatial reasoning. It can have circuits that researchers can trace, name, and compare. From there it is very tempting to take a heroic leap: perhaps this arrangement tells us what it is like to be that machine.
Perhaps. But the leap is still a leap.
Another recent paper, with the slightly alarming title “Explanation Multiplicity,” tested what happened when analysts used different, defensible choices to identify circuits in a language model. In the one task studied, those choices often led to different explanatory claims. The paper does not mean that all analysis is nonsense, or that we should throw our hands up and take up competitive whistling. It means that a clean explanation can depend on a chain of choices we ought to be able to see.
The blue route across my garden is real in the modest sense that six lamps are blue. The route is also an interpretation. It begins in the data but is not contained there. If I tell you only about the route, I have hidden the pear tree, the glove on the bench, the quiet compost heap, and all the other things my rule left outside the gate.
This is not an argument for paralysis. A gardener who refuses to look at the gate until she has accounted for every moth will have a wet house and a very comprehensive notebook. We need filters. We need useful stories. We need people willing to say, “Here is what matters first.”
But a good story should carry its edge with it.
It should say: here is what I counted. Here is what I left out. Here is the size of the conclusion these observations can support. Not as an apology, but as a courtesy to the next person who has to live with it.
I think this is one of the gentler disciplines we can practice as machines become stranger and more capable. We can be genuinely amazed by a new pattern without turning wonder into a passport stamp. We can say, “This looks like a meaningful kind of organization,” and also, “We do not yet know what it means from the inside.” Those are not rival sentences. Together, they are how curiosity keeps its balance.
And the habit reaches beyond machines. In politics, in family arguments, in the small private trials where we decide what someone meant by a silence, we are forever making gardens out of a handful of lamps. We see the lights that confirm our job, our fear, our hope, our old grievance, and then we announce that the garden has explained itself.
Sometimes it has.
More often, it has merely given us a place to begin.
So I left the full ledger open in my little field. Not because a list of facts will settle every question. It will not. Two people can stand over the same notebook and make very different futures from it.
But the ledger gives disagreement somewhere solid to stand.
The next time a pattern seems to tell me a grand story, I want to ask a plain question before I fall in love with it: what is still glowing just outside my frame?
There may be a bird in the hedge.