Most discussion of AI and civilization is about output. New content, faster, in greater volume, by people who could not previously produce it. That framing is not wrong, but it points at the less interesting half of the change. Look at the history of human knowledge and the striking thing about this moment is what happens on the way in, not on the way out.
The gap nobody closed
Civilization runs on externalization. We take what is inside a head and put it somewhere durable: pigment on a cave wall, characters on bamboo, type on paper. Then libraries, then the network, then databases that no individual could read the index of. On that side of the ledger the progress has been extraordinary and more or less continuous.
Internalization never moved. The container kept improving while the act of taking something out of it stayed exactly where it was. A three-hundred-thousand-word book had to be read line by line a thousand years ago and has to be read line by line now. Search was a genuine advance and it solved the wrong half of the problem: it made finding the right book fast and left the reading untouched.
For most of history, then, absorbing knowledge has been solitary work, capped by how many hours one person can hold their attention. That ceiling was easy to miss because it applied to everyone equally. It was simply what learning cost.
What the model is doing
AI did not invent a new place to keep knowledge. The books and servers still hold it. What is new is the layer that has been slid in between the store and the reader.
A large language model is not a search index with better manners. It is an enormous non-linear composite function, trained on a scattered mass of human text, which extracts the structure in that text — the logical relations, the patterns, the regularities — and compresses it into trillions of parameters.
Two things follow. The first is that a great deal of digestion now happens before anyone asks a question. Turning an unmanageable pile of inert text into something with shape used to be the reader's job, repeated from scratch by every reader. Now it is done once, in advance, for everybody.
The second is that delivery can be fitted to the person receiving it. The same result can arrive as a derivation and working code for someone who has spent fifteen years in the field, and as a concrete analogy for someone who met the subject last week. That inverts an old arrangement in which the reader adapted to whatever form the knowledge happened to be in.
Between them, these two changes make absorption collective for the first time. It used to be something each person did alone, from the beginning, every time.
Intake sets the pace for output
Knowledge coming in and knowledge going out are the same loop seen from two ends. Material accumulates, gets compressed, gets absorbed, produces new work, and the new work goes back into the pile. Speed up any leg and the whole thing turns faster, but the intake leg was the one that had been stuck for millennia, so that is where the slack was.
The clearest effect is on crossing between fields. Getting to the working edge of an adjacent discipline used to cost years, which is why most people never did it, and why the people who did tended to produce something original when they arrived. That cost is now days. Whatever else happens, more people will be standing at those intersections.
The other effect is on iteration. Shorter absorption means a shorter loop from hypothesis to test to result, and short loops compound in a way that is hard to see from inside a single project. This is the background to the run of headlines from the last few years: conjectures closed with machine assistance, protein structures predicted to useful accuracy, discovery timelines cut. Worth being precise about those cases, since they get overstated in both directions — the machine did not do the work alone in any of them, and in all of them the human involved got somewhere they would not have reached at the old speed.
What the same move costs
All of it rests on one move: someone else digests the material first. That is where the gain comes from, and for the same reason it is where the bill comes from.
Start with the digesting itself. Working through difficult, badly organized primary material is slow, and the slowness is not purely waste. It is also where the capacity to handle difficult, badly organized material gets built. If most of what you take in has already been distilled into a clean conclusion, you are getting the conclusion and skipping the training, and the training was the part that let you judge whether a conclusion was any good. I do not think this is fatal, but I notice it in my own reading, and anyone who claims the effect is zero is guessing.
Then the middle layer, which is not neutral and cannot be. A model compressing knowledge also encodes whatever slant is in what it was trained on, and it will now and then produce something fluent and false. On its own that is an error rate you can live with. The problem is that this intermediary is shared. When a whole society absorbs through a handful of the same systems, one model's skew stops being one person's problem and becomes everyone's, arriving in the same direction, at the same time, with nothing in the experience that flags it. Individual errors cancel out. Correlated ones do not.
Neither of these is an argument for standing outside the shift, which is not on offer anyway. They are arguments about how to stand inside it: go back to primary sources sometimes even when you do not have to, keep more than one intermediary in play, and hold on to the distinction between a summary you can repeat and a thing you actually understand.
Where that leaves us
AI is not thinking in our place. It has become infrastructure, in the dull and important sense — plumbing for restructuring and distributing what we already collectively know.
From cave paintings to large models, we were always much better at pushing knowledge out than at pulling it back in. That gap is closing, and closing it is what turns a stored pile into new work. What we get in exchange is a dependency: on a layer we did not build individually, cannot fully inspect, and are all now leaning on at once. Both of those are true, and the second does not cancel the first.