Attention isn't a flashlight you hold — it's a race for a scarce slot inside your head, and you only get to press one thumb on the scale.
Right now you're reading this line — and at the same time the pressure of the chair under you, the room's background noise, a whole clutter in your peripheral vision are all there, yet you barely "feel" them. Last issue we said perception is the brain guessing at the world; this issue's question goes one step further: the brain guesses at so much — why is it just "these few things" that made it into your consciousness while everything else got thrown out? The answer is attention. And once you see how it works, you'll notice something uncomfortable: the part you can actively control is far smaller than you think.
First, dismantle the most common metaphor: attention is like a flashlight — wherever you point it, that spot brightens. Sounds right, but mechanistically it's backwards. What actually happens: the neurons representing different things in your brain are constantly fighting each other — fighting for limited downstream processing, for the right to make you respond. They compete by default; what attention does isn't "light one up," it's quietly "tip the odds" toward one, making it more likely to win that contest. In neuroscience this is called biased competition.
That "contest + a thumb on the scale" picture instantly explains the famous cocktail party effect: a room roaring with voices, yet you can lock onto one person and follow their words — that's your goal "weighting" that voice so it beats the others. But while you're focused, someone across the room says your name and your head snaps around — you weren't listening over there, yet the stimulus "your own name" carries built-in weight and forcibly wins the contest. So the race is part yours to set, part completely out of your hands.
You've probably heard that the core of what runs models like ChatGPT is called "attention." The name collides, which makes it easy to assume they're the same thing — but they only coincide at the most abstract level: both "weight information by relevance and pass along only what matters." Underneath, the mechanisms differ sharply. A Transformer's attention takes each word, generates a set of query / key / value vectors, computes relevance by pairwise dot products, and sums everything in one parallel weighted pass (essentially one big all-to-all table of "who is most relevant to whom"). The brain has no such table; its attention is fought out bit by bit through neurons competing, plus the thalamic gate and gain modulation coming up. Hold onto this: functionally they're cousins, mechanistically they're two different things — a distinction you'll need every time a later issue hits a "biological X ↔ same-named AI X."
We just said the contest is part yours, part not — and that maps neatly onto attention's two sources, pointing in opposite directions:
Bottom-up (you get grabbed): some stimulus is so glaring it drags your attention over whether you like it or not. A sudden flick in the corner of your eye, a "bang" behind you, a red dot popping in the corner of a screen — before you can decide, your head and eyes have already turned. This is run by the stimulus's own "salience," hardwired by evolution: things that move, change, might be dangerous get priority.
Top-down (you send it out): you cast attention with a goal. Hunting for your red car in a parking lot, suddenly every "red" in the place jumps out; switch to hunting a license plate and now the digits jump out instead. Same world — you've fitted it with a different filter as needed.
The rough part is that these two fight, and the barging-in one often wins. You resolve to focus on writing (top-down), your phone goes "ding" (bottom-up), and your attention is hijacked — the entire ad industry, every app's red dots and push notifications, is in one single business: manufacturing salient stimuli engineered to overpower your goals. You're not weak-willed; you're up against a machine optimized to grab you.
Why does the brain go to all this trouble to "select"? Because downstream processing is a bottleneck — at any one instant, the amount of information that can truly be deeply processed and pushed into consciousness is pitifully small. A blur of world pours in, the gate is only so wide, and the vast majority is thrown out before you're ever aware of it.
The most famous proof is the "invisible gorilla" experiment: you're asked to count how many times players pass a ball, and you count in fierce concentration — meanwhile a person in a gorilla suit strolls right through center frame, even thumps their chest, and about half of viewers don't see it at all. It's not that the eye failed to capture it — the light landed clearly on the retina; it's that it didn't win the contest for attention, so it never entered consciousness. This is called inattentional blindness.
The same bottleneck bursts the myth of "multitasking": you think you're doing two mentally demanding things at once, but the brain is actually switching back and forth fast, paying a "reload" tax on each switch — so texting while driving isn't divided attention, it's a few seconds every so often when no one is minding the wheel. Real parallelism only happens when one of the tasks has been drilled to the point of needing no thought (like walking).
"Attention is scarce and costs something" has a sharp echo on the AI side. A Transformer has every word "attend to" every other word, and the compute blows up quadratically with length — to have a model read a context as long as a whole book, the attention term alone gets insanely expensive. So a whole line of recent research is doing the same one thing: don't let it attend to everything, pick only the relevant handful (sparse attention, all sorts of "only look local / only look at the gist" approximations). Full circle: pushed by compute, artificial systems rediscovered what the biological brain was forced into by energy from the start — attending to everything is too expensive, you must select.
After all this talk of "select" and "filter," where exactly does the gate sit? One crucial spot is a cluster of nuclei dead center in the brain called the thalamus. It's the master relay station to the cortex (that outer "shell" doing the high-level processing): except for smell, nearly all your senses — sight, hearing, touch — have to change trains in the thalamus before reaching the cortex. Thalamus
The clever part is that the thalamus is wrapped in a thin layer of neurons dedicated to "hitting the brakes," called the thalamic reticular nucleus. It sends no information to the cortex itself; it does one job: inhibit the relay cells inside the thalamus — that is, decide "which channel of signal gets through, which gets cut." Nobel laureate Crick likened it to the brain's searchlight gate: the channels attention wants to suppress may be exactly the ones quietly turned down right here.
There's an even more striking corollary: this gate doesn't only tune "where you attend," it tunes "whether you're awake at all." In deep sleep and under anesthesia, this thalamus-to-cortex route is largely shut off, sensory signals can't get up — and that's part of the mechanism of your "losing consciousness." In other words, the same gating system, dialed fine is attention, dialed shut is falling asleep. How consciousness gets switched is the big show of Phase B later; this gate is its earliest foreshadowing.
"A dedicated little module that does no work itself, only decides which channel of signal gets through" has a rather apt cousin in AI called gating / routing. Take the fashionable mixture-of-experts architecture: the model keeps a bunch of "expert" sub-networks, with a gating network out front that processes no content itself and only judges "which experts should this input go to, gate the rest off." The thalamic reticular nucleus is to the cortex a bit like that gating network is to the experts — both are a dedicated valve that hands precious compute only to the channel that should get it.
"Where the mind is placed, that's where the world takes shape" — several old traditions have long treated attention as a first-rank matter, to be trained and pondered: