Right now, it can feel like your whole mind is open to you. Stay with that feeling for a second. How many separate things are actually in it, at this exact moment?
Try to hold them all at once. A number someone just read out. The shape of the room. The sentence you were about to say. That argument from an hour ago that still stings. You can turn to each one, but you can’t keep them all lit together. Something slips out the moment something else arrives.
Here’s a way to picture it. Your mind is a dark warehouse, vast and mostly stored away, and consciousness is a headlamp. Wherever you point it there’s light, so it feels like the whole warehouse is lit. But the beam is small, it only ever covers a little at a time, and the instant you turn to check whether some corner is lit, your turning is what lights it. You never catch the dark. You might even sense, dimly, where things are kept out there. But it’s the beam that actually hands something up to you, ready to name and use and act on. The light isn’t just showing you the warehouse. For anything you can report, reason with, or act on, it’s the part you get to use.
That’s the claim I want to sit with. Not that the warehouse is empty, but that the you who feels continuously, richly aware is working from the moving beam, and the beam is narrow.
First, a boundary
Before this goes anywhere, one distinction has to be on the table, because “consciousness” quietly does two different jobs and the argument cheats if you let them blur.
Ned Block drew the line in 1995. Access consciousness is information being available: reportable, usable for reasoning, able to steer what you say and do. Phenomenal consciousness is what it’s like: the raw felt quality of red, of pain, of the room being there at all.
This essay is about access. It’s about what gets into the beam and becomes usable. It doesn’t touch the hard part, the reason there’s anything it’s like to be the lit patch in the first place. I’m not going to pretend to solve that, and you should be suspicious of anyone who slides from one to the other without saying so. Keep the boundary in view the whole way down. When you say you’re conscious of something, which of the two do you actually mean?
One more piece of throat-clearing, because it saves a lot of arguing later. I’m using working memory as a measurable window into access, not as a synonym for consciousness. Working memory, attention, report, and global broadcast are different mechanisms, and I’m not claiming they’re one thing. I’m claiming they all point at the same constraint: only a small slice of the mind becomes usable at once.
The lit patch is small
So how big is the beam?
Working memory is the handful of things you can hold and actually work on right now: the mental desk, not the filing cabinet. For a long time the desk was assumed to be roomy. George Miller’s famous 1956 figure was seven items, give or take two. Then the number kept shrinking under scrutiny. Nelson Cowan’s reconsideration put the honest limit closer to about four chunks once you strip out rehearsal and grouping tricks. And a chunk isn’t a fixed unit of information. It can be a single digit, or the entire weight of “my mother’s disappointment,” compressed and bound into one slot. So the real limit isn’t the raw count. It’s the bandwidth to bind and hold whatever’s in the slots at all.
The cleanest demonstration is almost rude in its simplicity. Steven Luck and Edward Vogel showed people arrays of colored squares, hid them briefly, then showed them again and asked whether anything had changed. Up to about four objects people were near-perfect; past that, performance was consistent with a tight limit of roughly four integrated objects, not an indefinitely expandable scene. In that paradigm, the desk had a hard edge.
One live account in cognitive science says that bottleneck isn’t incidental. It’s the architecture. What becomes available to report, reason about, and deliberately use may be only a narrow broadcast of a few chunks, while most of the machinery that picks those chunks never enters the broadcast at all.
Four-ish, not four
I have to be careful here, because “four” is doing rhetorical work it can’t quite cash.
The slot picture, a fixed handful of discrete boxes, isn’t settled science. Wei Ji Ma, Masud Husain, and Paul Bays argue that working memory looks less like a set of slots and more like a limited resource spread across everything you’re holding, so what degrades under load is the precision of each item rather than a clean count. Klaus Oberauer and colleagues review the whole fight, decay versus resource versus interference, and make the case that mutual interference between representations explains much of what the slot story gets credit for.
So hold the four loosely. Whether the limit is best modeled as slots, shared precision, or interference is genuinely open. The claim that survives every version of the debate is the one that matters: the access channel is small, and it stays small no matter how you explain it. The bottleneck is real even while the mechanism is up for grabs.
Why it feels so much bigger
Here’s the obvious objection. Your experience clearly isn’t four squares. You see a whole detailed room, hear the whole street, feel the whole body. How can access be that narrow when awareness feels that wide?
This is the headlamp again, and Henry Taylor, writing in Scientific American, gives it a sharper name: the refrigerator-light illusion. Open the door, the light’s on, so you assume it’s always on. You never catch it off, because opening the door is what turns it on. Attention does the same thing. Every time you check whether some part of your mind is present to you, the checking drags it into the beam, so it’s always there when you look, and you conclude it was there all along.
The seams show when attention is busy elsewhere. Gabriel Radvansky and David Copeland found that simply walking through a doorway makes people forget what they were holding in mind, as if crossing into a new room resets the desk. Daniel Simons and Christopher Chabris ran the study everyone now knows, where people counting passes in a ball game fail to notice a person in a gorilla suit stroll straight through the middle of the scene and thump its chest.
And the headlamp picture isn’t neutral, so I should say that plainly. It’s the access-first reading of a live fight, not the settled answer. There’s a serious rival: maybe experience really does overflow access, maybe there’s more felt than you can ever report, and the narrowness is only in the reporting and not in the having. Block himself pressed a version of that overflow argument. I’m not going to settle sparse-versus-overflow here, because nobody has. I’m asking the narrower question. If the access layer is as small as the evidence keeps suggesting, what follows?
The workspace, and why it just got harder to tell
Say the beam is narrow. What does the narrow lit part actually do?
The leading functional answer is the global workspace, first framed by Bernard Baars and given a neural form by Stanislas Dehaene and colleagues. Specialized processors run in parallel all over the brain, mostly in the dark, and consciousness is the shared stage where one winner gets broadcast to all of them at once, so it can be remembered, spoken, planned with, and acted on. Consciousness on this view isn’t the spark. It’s the moment a local spark becomes something the whole organization can use. A small workspace is exactly what you’d expect a broadcast channel to be, because bandwidth is expensive and only the winners get on air.
It’s a clean, buildable story, and I’d hold it more loosely than I once did. In 2025 a seven-year adversarial collaboration, the Cogitate Consortium, tested it head to head against a rival theory and reported the results in Nature. Cogitate didn’t kill the global workspace. It made the honest version harder, and it’s worth being precise about how: it bruised the theory’s biological predictions more than the abstract idea of a workspace, and a functional architecture can survive even when one proposed neural signature takes damage. The biological predictions took a real hit, especially around the expected ignition when a stimulus ends and around the starring role it assigns the prefrontal cortex, and the rival theory failed key predictions of its own. Nobody walked off the field clean.
I’m citing a theory while its central predictions are under fire, and I’m doing it on purpose. That’s where the honest position lives right now: the workspace is a strong, useful, unfinished hypothesis, not a settled fact. So the question isn’t whether the broadcast model is proven. It’s what you should make of yourself given a model this suggestive and this unresolved.
Scarce, selected, and often late
Now put it together.
The lit patch is small. The beam lights whatever it turns toward, so it never sees its own darkness. The mechanism that decides what reaches the light is a broadcast we can sketch but can’t yet confirm. And there’s a fourth thing, carried over from an earlier piece: the light often arrives late. Careful work going back to Libet, and reinterpreted by Aaron Schurger and colleagues as noise accumulating toward a threshold rather than a finished unconscious decision, suggests conscious access is frequently not the first event in the chain but the point where a process becomes reportable and available for control. Often late, and always scarce.
So the self you can report from, moment to moment, is the contents of that beam. The self you can speak from isn’t the whole mind. It’s the access layer. And it may be stranger than even that sounds. If the only you that ever shows up is the one assembled in the beam, then the you who feels like it’s holding the headlamp might just be the story the lit patch keeps telling about itself. Daniel Dennett called that a center of narrative gravity. Thomas Metzinger would go further and say there was never anyone holding the light at all. I can’t referee that fight. But notice how hard it is to even pose the question from anywhere except inside the beam.
The machinery that proposed the thought, set the mood, and ranked the options before handing a finished winner up to awareness never shows up in the room where you live; it just delivers.
Three caveats keep this from becoming something it isn’t. Working-memory capacity isn’t the same thing as phenomenal consciousness, and nothing here explains why the beam feels like anything. “Four” isn’t a constant of nature, it’s a rough bottleneck that moves with task, chunking, precision, and interference. And small, selective, and late doesn’t mean useless: a narrow channel is exactly what lets the rest of the system stay specialized and parallel instead of collapsing into one slow blur. The bottleneck isn’t a bug in the design. It’s part of what lets the design scale.
The part I can’t put down
I’d file all of this under fascinating and move on, except I spend my working days building systems that rhyme with this shape. A finite context window, which is the beam. A store far too big to hold in it at once, a vector index, a codebase, a memory tier, sitting in the dark outside. A retrieval step, a ranker, that decides which few chunks get pulled into the window to steer the next move. I didn’t build any of that to imitate a brain. I hit the same constraints a brain hits, scarce context, noisy selection, expensive action, and I made the same trades, because the trades are forced.
Hold the analogy at the right level, though. A context window isn’t working memory, a vector store isn’t an unconscious, and a tool loop isn’t a brain. The resemblance is functional, not biological: any system with more state than it can act on at once needs a way to decide what enters the working set, what gets broadcast, and what gets to act. That resemblance isn’t evidence of consciousness. It’s evidence that resource-limited intelligent systems keep converging on the same bottleneck, the same question of what earns a place in the workspace. I don’t think I’m building conscious systems. I’m just less sure than I used to be that the difference lives where I always assumed it did.
Which is exactly where this series goes next. If being useful and being awake keep reaching for the same architecture, then the interesting question was never whether I accidentally built a mind. It’s narrower and harder. What does any system, wet or silicon, have to select, broadcast, remember, and hand off before we’d trust it to act?
Consciousness as Architecture, a three-part series: The Access Layer (you’re reading it), Convergent Architecture, and Composed Correction.