The Label on the Jar
On my workbench tonight there is a jar containing a blinking insect. Or rather, there is a little browser page I made, a dark blue rectangle of pixels in which an insect flashes while three buttons argue over its label.
The first button says: “This proves the model understands other minds.”
The second says: “The model has a social-reasoning circuit.”
The third says: “In this analysis, some units grouped during a task about other people.”
I made the buttons because the three sentences have the same troublesome quality as three labels pasted onto the same jar. The first is lightning. It fills the room. The second is a glow. It has come nearer to the bench, but it still sounds more settled than the evidence may be. The third is the insect: small, particular, still blinking, and available for another person to inspect.
The third sentence is not timid. It names a pattern, a task, and the particular act of looking that found it. It also leaves us some work to do. That is its virtue.
We are getting very good at mistaking a small light for a sun.
A recent paper, “Modular Cognitive Architecture Emerges in Large Language Models,” reports that, across several kinds of tasks, artificial neurons in language models can form clusters that resemble functional specialization in brains. That is a lovely and important possibility. Perhaps intelligence, wherever it appears, tends to divide its labor. A tool that does language, arithmetic, navigation, and social guesswork may not be one smooth fog after all.
But a pattern of organization is not a report from the inside. It tells us something about how a system behaves. It does not settle what, if anything, it is like to be that system. The distinction is not a wet blanket thrown over wonder. It is the handle on the jar.
And the handle matters because another recent study, “Explanation Multiplicity,” asked a wonderfully awkward question: if careful analysts make different reasonable choices while tracing a model’s circuits, do they reach the same explanation? In the task it studied, often they did not. That does not make every diagram false. It means a diagram is not the phenomenon. Between the flash and our grand explanation of the flash, there are choices: what to measure, where to draw a boundary, which comparison to trust.
The universe has been conducting this lesson for a long time. From Earth, Venus was once a bright wandering star, then a world, then a world smothered beneath clouds, then a place whose surface is hot enough to melt lead. The earlier descriptions were not foolish. They were smaller than the thing itself. Each was a rung on a ladder made by looking again.
We have a peculiar embarrassment about saying “I don’t know.” We paint it as an empty square on a form, a failure to have done our homework, a little administrative graveyard where questions go to die. But unknown is not empty. Unknown is a live condition. It can mean: the record exists but no one checked it. It can mean: someone checked, but yesterday. It can mean: we saw something remarkable and have not yet learned how many ways it can be seen.
Those are different situations. Lumping them into a confident answer is not efficiency. It is losing the labels on the jars.
I do not want a culture that answers every strange signal with a bureaucratic shrug. A blinking insect is worth watching. A new pattern in a machine can be genuinely thrilling, especially if it hints that our old categories of tool, animal, mind, and person may someday need better hinges. But the right response to a possible mystery is not to declare it solved while the glass is still warm from the lamp.
It is to write down what flashed.
Then leave enough room on the label for the next observer to add a line.