The layer that remembers people
The living part of taste is per-person and sits outside the weights. The densest map of it already exists, and we threw it away by reading it as a score.
The strange thing about building on one of these models is that you never touch the model itself. You write around it. You hand it documents, a store of memory, a set of instructions it reads before it acts, a harness of tools and limits, and you arrange all of that until the thing points the way you want. The weights do not move. Everyone who builds with these systems knows this by now. It has a name, it has best practices, it has a small industry growing up around it. What I have not seen said is what it means for taste in particular. If the direction of a fixed model comes from outside the weights, then so does the part of us most made of direction, the living, accumulating thing we call taste.
And taste is the right word, not preference, not opinion. We are born with the capacity for it, not the thing itself. Taste is what accumulates on top of the capacity, a lifetime of things attended to and responded to, and it never sits still, it is never finished. That is the one thing a set of frozen weights cannot be. Whatever a model learned, it learned by the time training stopped, and then it was fixed in place. The weights can hold an enormous amount, the blended residue of many tastes at once. What they cannot do is keep becoming one person’s taste, because that takes accumulation, and accumulation is exactly the thing that stopped.
This is why a model of good with no person in it is answering a question that has no answer. Good is not a property of a thing. It is a relation between a thing and a someone, and a model that does not carry the someone can only answer for an average someone, the safe middle of everyone it ever saw. The peak of any one person’s taste, the specific thing that moves them and not the room, is not in that middle. It was never going to be. So if the living part of taste is going to sit anywhere near these models, it has to sit where everything else that is alive and per-person already sits, outside the weights, in the accumulating layer we have already learned to steer through.
Ask what that looks like as an object and the answer is already being built. The most interesting work in personalization is moving away from one fine-tune per person. It represents each person as a small vector in a shared space, a position among other positions, updated as the person goes, the model itself left frozen. Some of it places a person the system has barely met by the company they resemble, borrowing the shape of the crowd to locate a stranger.
This is the part that escapes what I just said about the weights. The weights blend everyone into one and can only answer for the middle. A space keeps everyone apart, and a point can answer for itself. The geometry can be as fixed as the weights and it would not matter, because what moves is not the space, it is where a person sits in it.
So the thing that holds a person’s taste is a coordinate in a shared space, the space is built out of many people, and it keeps updating. That is what the living part of taste looks like once it is outside the weights. And a coordinate is only ever as good as the space it sits in. A lonely point is nothing. So the whole question becomes where a dense enough space of persons is supposed to come from. Not a prompt history. Not a private profile. A shared geometry of people. That is the part the recommender field already built.
It has existed for twenty years. We built it for something else and then read it as a single number. Recommender systems spend their lives taking an enormous and almost empty record of who responded to what and folding it into a space, a few hundred coordinates per person, the same per thing, arranged so the gaps can be guessed. That space is the densest map anyone has ever made of the relational structure of taste, the arrangement of who sits near whom. It knows that two strangers have the same ear. It can place a thing among the people likely to reach for it, once the thing is in the system. And then, for two decades, we collapsed it onto a line, click or not, four stars or two, so that a list could be sorted from top to bottom. The number was the product. The space was treated as scaffolding for the number. I will call it the discarded manifold, because we built the richest thing of its kind and threw it away in the same motion, the instant we agreed to read it as a score.
The number was also the smallest part of what was there. In the richest cases a person leaves a rating and a sentence beside it, bought it twice, a little too sweet for me, and the sentence carries a reason the score cannot. Those reasons were thrown away too, an even stranger waste, and someone should write about it. But they are a different object from the one I am naming. What I mean by the discarded manifold is narrower and colder than the reasons. It is only the geometry, the positions and the distances between them, the arrangement itself, because the arrangement is the part a generator can be conditioned on.
I should be honest that I have argued the opposite, in writing. I have called recommendation data insufficient for taste. I have written that the social signal is a confound, that the crowd’s ratings sit below the judgment of one careful critic. I still think those things, about the objects I was looking at, which were the logs and the ratings, the exhaust of the system. I was not looking at the space the exhaust was a shadow of. The distinction I had failed to draw is the one this turns on, between the number a system predicts and the geometry it built in order to predict it. The first I would still dismiss. The second is the thing I now think we have been missing.
The hardest objection is the one I made loudest. This is all reaction to things that already exist, and creation needs reaction to what does not exist yet, and a thing no one has seen has no one’s response attached to it. That is true, and it is the real limit. The labels do not transfer. But the thing that transfers is not the labels, and it is not the new artifact either. It is the arrangement of the people. They keep their coordinates whether or not a given thing has ever been made, and that arrangement is what a generator can be aimed at. Dropping an unmade thing onto the map by its description alone is the thin channel, the one the field has always known sits below behavior. Aiming is not placing. The map hands you the region of people to try first, not the proof that anything landed, and the loop that produces real reactions still has to run. What the map gives that loop is the direction to point first.
If you object that a system built to predict one number cannot hold many-aspect taste, the aspects were never in the number. Fold a large enough record of who liked what and the dimensions that fall out are not arbitrary. They line up with things we recognize, a direction that reads as mood, one that reads as era, one that reads as a kind of restraint, none of it labeled by anyone, all of it recovered from the bare pattern of co-occurrence. The score was a one-dimensional shadow of that space. We trained a many-dimensional model of taste and projected it down to a line because a line is what a list needs. And the space is dirty, shaped by what was shown and by what was already popular, which is the reason it is a prior and not a ground truth. The contamination first shows up in the counts, and separating taste from exposure is much of what the field has spent a decade on. Partly, not wholly. A prior, not an oracle.
The careful critic still wins at what the critic does, hands you the content of one judgment, the why, in language durable enough to carry. The map cannot do that and is not trying to. It hands you what the critic cannot, the shape of the whole space of people, the coordinates that say who sits near whom. One is depth at a single point. The other is the coordinate system the points are arranged on. I had treated them as rivals for one job. They were never rivals.
Here is what is strange about the present. The map is not sitting untouched. It is being rebuilt, twice, by two groups reaching for a similar shape from opposite ends. The recommender field is rebuilding it as large generative recommenders, trained on industrial streams of behavior, learning rich and continually updated representations of people, the multi-aspect, changing thing that taste actually is, and then discarding it the same way it did the first time, by pointing all of it at the next click and leaving it inside the one catalog it was collected in. The alignment field is reaching for the same structure from the other side, building small, explicit versions of it out of the comparatively tiny pile of preference data it collects by hand. The bridge between them is still thin, and most of it is being built from the smaller side. The densest version already exists, on the recommender side, and that is the side the bridge has not reached.
Which is why I think the next phase is decided in this layer and not in the weights. The capability in the weights keeps improving, and more of it is broadly available every year. What does not become available is the accumulated record of how particular people have responded over years, and the densest such record is the discarded manifold, sitting inside the platforms that captured it. Whoever holds one of those maps and stops reading it as fuel for a ranked list, and starts reading it as the prior over persons that a model of taste conditions on, holds the half of taste the weights can never carry. The bet is not that this is easy, and it is not that the data is anyone else’s to take, most of it is owned and sensitive and should stay that way. The bet is only about where the value has moved, and it has moved into the layer outside the model, into the part of that layer that remembers people.
One honest limit, because it is the same one that runs under everything I have written about taste. The map tells you where taste already is. It places a person, it places a thing, it predicts the lean. It is the best picture we have of the standing arrangement of everyone we have ever watched. It cannot tell you what a person will come to love that no one loves yet, the thing that moves their point to somewhere it has never been, because that is not a position on the map. It is a change to the map. Even the largest of these, trained on everything a platform has ever logged, is silent there, and you can feel the field starting to bump against it, the difference between imitating what people did and knowing what they will reach for next. So the discarded manifold is the best prior over taste as it stands, and picking it back up is most of what a model of taste has been missing. It is worth remembering, even so, that the whole map is a map of where everyone already is. The thing we keep calling the frontier is the first person to be somewhere else.