*On why the feed should be a city, not a column — and what that does to the social network.*
Every so often someone tries to free the feed from its column. Instead of an endless vertical scroll, they give you a 2D plane you can push in any direction — an infinite canvas, posts scattered across it like notes on a wall. It always looks liberating in the demo. And it almost always fails the same way: within thirty seconds you are lost. You pan, you zoom, you find something, you pan again to find the next thing, and now you cannot get back to the first. The freedom becomes vertigo.
It is tempting to blame the second dimension. That is the wrong lesson. The problem is not that the canvas is 2D. The problem is that it is *blank*.
Disorientation is an absence, not a dimension
A blank plane gives you nothing to hold on to. Everything looks like everything else; nowhere is anywhere in particular; a position on the canvas *means* nothing, so there is nothing to remember. You cannot form a mental map of a space that has no features, for the same reason you cannot memorize the layout of an empty white room. Disorientation is not caused by having two axes to move along. It is caused by a homogeneous space with no differentiation, no persistence, and no meaning attached to where things sit.
Which is a clue, because we already have a class of 2D interfaces that people navigate all day without getting lost: maps. Nobody opens a map and panics. You pan, you zoom, you wander to the far edge of the city and back, and you always know roughly where you are. A map is a 2D canvas you can scroll in any direction — the exact thing that fails as a blank feed — and it works. So the interesting question is not "is 2D too much?" It is "what does a map have that a blank canvas doesn't?"
The Image of the City
In 1960 the urban planner Kevin Lynch published a small, strange, enduring book called *The Image of the City*. He was trying to answer a deceptively simple question: what makes some cities easy to find your way around, and others a fog? He interviewed residents, asked them to draw their cities from memory, and found that legible places are built from five kinds of thing:
- Paths — the routes you move along.
- Edges — boundaries that separate one area from another.
- Districts — regions with a shared character you recognize as "a part" of the place.
- Nodes — junctions and hubs you pass through.
- Landmarks — distinctive, memorable points you orient yourself against.
A blank canvas has none of these. A map has all of them. That is the entire difference. Legibility is not a mysterious property of good design; it is a checklist, and every navigable 2D surface — a city, a map, or a feed that wants to be either — has to earn each item.
Underneath the five elements sits one principle that does most of the work: position has to carry meaning, and it has to stay put. On a map, being *here* rather than *there* tells you something real — this is north of the river, that is the old town. And the river does not move overnight. A canvas becomes navigable exactly when its coordinates mean something and its geography persists long enough for you to remember it.
The trouble with mapping a feed
Here is where a content-map is harder than a real map, and where most attempts quietly break.
Geography is fixed. London is where it was yesterday, so the mental map you built last week still works today. A feed is the opposite of fixed — new things arrive constantly. If posts *rearrange* every time something new comes in, your spatial memory is destroyed on every refresh. You are back in the blank white room, except now the furniture teleports.
So the real design question is not "how do I lay the posts out." It is: how do I give this territory a stable geography that survives new content? How do I make a map whose coastline doesn't redraw itself every hour?
The whiteboard tools — the Figmas and Miros of the world — answer this by making *you* do the work. You place each object by hand, and it stays where you dropped it. That gives persistence, but the burden is entirely on the human, and it does not scale past a few dozen items before your own placements become their own kind of clutter. It is a workshop, not a world.
There is a better answer, and it is almost embarrassingly clean: don't place things by hand, and don't place them by arrival time. Place them by meaning.
Meaning is the stable geography
Modern language models turn a piece of text into a point in a high-dimensional space where *nearness equals similarity in meaning*. Project that space down to two dimensions and you get a map that draws itself: posts about the same thing land near each other, unrelated posts land far apart, and clusters of shared subject matter form — automatically — into regions.
This is the move that makes a content-map legible where a blank canvas cannot be, because it fixes the persistence problem at the root. In a whiteboard, position is arbitrary and manual, so it is fragile. In a semantic map, position is computed from what the content is about — and meaning is stable. The region for a given topic is always in the same place, because it is *defined* by meaning, not by when something showed up or where a hand dropped it. A new post about that topic simply lands in its region and raises the local density. It does not shuffle the map. The coastline holds.
And once the geography is stable, Lynch's five elements fall out of ordinary map conventions you already know how to build:
- Semantic zoom, exactly like zooming a real map: far out, you see district labels and a heat-map of where the activity is; mid-range, cluster titles; up close, the full posts. Detail appears as you descend.
- Clustering, like map pins collapsing into a numbered dot when you zoom out of a busy neighborhood.
- Districts are the topic regions; edges are the quiet gaps between them; paths are the threads that link related posts; landmarks are the standout posts and people; nodes are the hubs where many paths meet.
The content becomes its own cartography. You are not decorating a blank plane. You are reading a territory that the writing itself laid down.
The social turn: one city, many maps
All of that is true for a single person's information space. But a social network is not one person, and this is where the metaphor stops being a nice visualization trick and becomes something structural.
Go back to Lynch, because he saw this too. He was careful to distinguish the shared city from each resident's image of the city. The terrain is one thing — there is a single London. But the mental map of a cab driver and the mental map of someone who never leaves their neighborhood are almost entirely different. They overlap only at the big public landmarks. Same city; different images.
A social network is exactly that, and it resolves into two layers.
The base map is shared and objective. The semantic geography is the same for everyone. A post about a particular subject sits in the same region of meaning for you as it does for me. This layer is the neutral terrain — nobody's opinion, just where things are.
The landmark layer is personal. What stands out to *you* — what renders large, what anchors your wayfinding, which paths are lit — is drawn from your relationships, your history, the people and things you have chosen to attend to. My landmarks are not your landmarks. We live in the same city and carry different maps of it.
And the part that makes this genuinely social is the thing you might expect to be a weakness: our maps barely overlap. The intersection between your landmarks and mine is small — a few mutual connections, a couple of shared interests, the occasional post we both anchored to. But that small intersection is not a bug. It is the entire social fabric. Overlap is where connection lives: *we both know this place.* And the vast non-overlap is where discovery lives: because the base map is complete and shared, you can wander into a district lit only by someone else's landmarks and find something you would never have surfaced on your own. The territory is fully navigable even where your own salience is zero.
Why this beats the two feeds we already have
Look at what this quietly fixes.
A broadcast feed — everyone sees the same thing — is a single shared image with no personalization. One map for a whole city of very different people.
An algorithmic feed — a machine guesses what you want — is a personal image with the terrain hidden. You are shown your bubble and nothing else, and, crucially, you cannot see the shape of what you are missing. There is no map, only a conveyor belt someone else controls.
The map model refuses that trade. It gives you shared reality and a personal path at the same time. Your landmarks guide you, but the whole city stays visible; you can always pan somewhere unfamiliar. There is no invisible bubble, because the geography is public. And there is no gatekeeper deciding what is prominent, because prominence is not a dial someone turns on you — it is simply your own relationships projected onto neutral ground. The platform draws the coastline. You bring the landmarks. Nobody in the middle gets to decide what deserves to be a mountain.
There is a lovely practical dividend, too. Because everyone shares the coordinates, people can talk about places even when their maps differ. I can tell you something is "over in the legal district, near the contracts cluster," and it means something to you — even though none of my landmarks are yours — the same way two strangers can agree on where "downtown" is in a city neither of them maps the same way. Shared geography makes shared language about place possible. That is the substrate of a conversation, and a conversation is the substrate of a network.
The reframe
So here is the whole idea in one line: one neutral shared geography, N personal landmark-overlays, and the network is the intersection structure between them.
The feed stops being a column someone ranks for you and becomes a city you learn your way around. The map is drawn by meaning, so it holds still long enough to remember. The terrain is public, so there is no bubble and no gatekeeper. The landmarks are yours, so it is genuinely your city. And the small, precious places where your map and mine overlap — those are not the noise. They are the point.
We spent a decade arguing about whether feeds should be chronological or algorithmic, as if those were the only two shapes a river could take. They were never the only two. The third one is older than software. It is a map, and it has been teaching people how to find their way home for as long as there have been cities.
End of Transmission.