we see in silence

So your eyes signal your brain when they see light? Well, no. It is literally the opposite. In the dark, the rods at the back of your eye fire the whole night, a steady report that nothing is happening, and when light lands on one it goes quiet. The quiet is the signal. The title is Hannah Fry's, from The Rest Is Science last week, an episode I only found because where the plant is had sent me reading about the senses. I would never have thought it worked that way, and now I keep wondering how many other things work backwards from what we assume. In the senses? In architecture? In the tech I build with?

Fig. 21 · A rod's output in the dark and in the light, at the back of the eye, drawn twice. Above, the way anyone would assume it works; below, the way it does. The quiet trace, in red, is the signal.

To be exact, because I got this wrong on the first listen: it is the rods that go quiet, not the brain. The cells behind them read the silence and pass it on, and by the time the message reaches the cortex it is a signal like any other. The backwards step is the first one, in the retina. And it is not a quirk of the retina. It is the current best guess about how the whole brain works: you do not receive the room, you guess it, top down, and the senses send back only the difference between the guess and what arrived. The expected thing produces no signal at all. The cells that hand out dopamine do the same, firing for the reward you did not see coming, silent for the one you did, and dipping when the one you expected never shows. Which is why you never hear the fridge until it stops.

It took me a week to notice that this is exactly what a language model does. It guesses the next word, and it turns out so do you: put electrodes on people listening to a podcast and the cortex predicts each word before it lands, then reports the surprise. In experiences are data I asked whether we had described one process twice in two vocabularies. Apparently we had. The difference is what happens after the guess. Mine is corrected every second by my own senses; the model was corrected once, in training, and never again. And I have a second way to cancel an error that no model has: I can get up and move the plant.

Now put a building in front of it. A facade the brain predicts perfectly, the same window forty times, produces no signal. So a boring building is not a matter of taste. It is literally less stimulating, because it was predicted. Colin Ellard walked people past a blank Whole Foods wall in Manhattan and their arousal dropped, and they said so, and Heatherwick's Humanise campaign has been arguing since 2023 that boring buildings are bad for us. This is at least a mechanism for why: nothing wrong, nothing to report, nothing happening in you. I want to be careful, because the opposite failure exists and every airport proves it. A room that surprises you at every step is not stimulating, it is exhausting. The old research on this has a shape, an inverted U, and comfort sits somewhere on the slope, not at either end.

Fig. 22 · Response against surprise, after Berlyne. Both ends sit low: a fully predicted room on the left, a chaotic one on the right. The hollow point marks a comfortable room, on the slope rather than at the peak.

Which leaves me with a drawing problem. Sensi scores a room across six senses, and right now a perfect score is a full bar, the way every score bar since scores began has been full. Maybe it should be the opposite. If comfort is the sense going quiet, the honest drawing is the empty one: the comfortable room is the one with nothing to report, and a full bar is a room that would not stop talking. The wrong thing is the thing you notice, I wrote in comfort as data, before I knew why. So do I redraw it? And would anyone trust a tool that shows them nothing?

Fig. 23 · Six sensory scores for one comfortable room, drawn two ways. Left, the convention: every bar full. Right, the same room drawn as what is still reporting: five silent, one talking, in red.