Amazon started with books. Now the company it became is feeding books into machines. That should probably make us think beyond the obvious story.
How often do I think about the Roman Empire?
Many do, of enough it was a trending question just a few years back.
But who thinks about Alexandria. That year I read a book about Cleopatra instead.
We have reached the part of the AI story where the company that started by selling books is buying physical books, cutting off the bindings, scanning them into machines, and destroying the originals.
There is nothing symbolic about that at all. There is, however, some historical context worth remembering.
404 Media recently tracked a number of books from a bulk purchase using AirTags. They eventually landed at an Amazon facility in Las Vegas where books were being stripped, scanned, digitized and discarded.
Anthropic has done the same thing at scale. Federal court records describe millions of physical books being purchased, destructively scanned and converted into a permanent digital library.
The immediate question is obvious: Why are AI companies buying old books and destroying them?
On the surface, the immediate answer is obvious: they need high-quality human-generated information.
But why destroy them? That is the deeper question. The next question, as my military friend always taunts me with.
What happens when the companies building the intelligence layer acquire the underlying record?
The Library is a Historical System
Just like in Alexandria.
The Library of Alexandria was never just a building full of scrolls. The larger project—the Mouseion and the intellectual system surrounding it—was an institutional of advancement of a society.
A pillar of information.
A gathering place of thinkers.
Funded by the society.
The library protected the continuous work of collection, comparison, correction, argument, discovery and iteration.
It created enough intellectual density and understanding compounds.
The names survived: Euclid, Eratosthenes and others.
The system didn’t.
Everybody remembers Rome, the Empire. Know one is going around asking people, how often do. you think about the library of Alexandria, but I do.
I am a little weird.
Underneath that historical contrast is a pattern that appears repeatedly.
Civilizations build capacity. They create institutions capable of accumulating knowledge, organizing people and solving difficult problems. Then they enjoy the results of that capacity.
Eventually, they can begin to stop protecting the conditions that produced it.
Prosperity turns toward extraction. Internal division grows. Institutions may remain standing long after the underlying generative strength begins to weaken.
And eventually something stronger, richer or better organized absorbs what remains.
Ray Dalio describes part of this through his Big Cycle framework. I tend to reduce the pattern further:
Build. Prosper. Fragment. Enclose. Absorb.
It is the enclosure part that interests me now.
You Don’t Have to Destroy the Library
Knowledge has always been enclosed.
States controlled maps and navigation information.
Governments classified technical knowledge.
Empires moved manuscripts, cultural records and artifacts.
Religious and political institutions established official canons and excluded competing interpretations.
You do not have to destroy information to control it.
Sometimes you simply move it behind a wall.
AI gives us a new version of that wall.
You can ingest the library.
And then build the machine everyone asks what was in it.
That changes the structure considerably.
Because now we are not simply concentrating information.
We are increasingly concentrating information + interpretation.
And information and interpretation are not the same thing.
The historical record is one thing.
What a model tells us about the historical record is another.
The distinction becomes increasingly important as the interface becomes easier than the source.
Most people are not going to leave an answer generated in two seconds, locate a 400-page book, find the original documents behind it and independently reconstruct the argument.
They will ask the machine.
And the machine—or more precisely the systems surrounding it—determines what gets retrieved, weighted, summarized, connected, ignored or left out.
It does not have to lie.
That is the part people miss.
Selection is power.
The New Information Asymmetry
No, I am not suggesting Jeff Bezos is sitting somewhere rewriting the Roman Empire.
That would actually be a much simpler problem.
The structural problem is larger.
A small number of companies are building systems increasingly positioned between human beings and the accumulated record of human knowledge.
At the same time, those companies are acquiring enormous quantities of the record itself and destroying the legacy.
That creates several forms of asymmetry at once:
Information asymmetry.
Interpretive asymmetry.
Decision asymmetry.
Narrative asymmetry.
Which brings this straight back to the larger AI economy.
We have spent plenty of time discussing who controls the chips.
Good.
We are finally beginning to discuss who controls the energy required to operate the infrastructure.
Also good.
But there is a third strategic resource sitting directly in front of us:
Human information.
Some of it is effectively finite.
Original documents. Books. Photographs. Research. Letters. First-person accounts. Historical records. Material that never made it onto the open web.
And increasingly important: material created before generative AI began producing enormous quantities of synthetic text that now circulates back into the information ecosystem.
The clean human record.
Once that material becomes strategically valuable, capital does what capital generally does.
It acquires it.
Once enough of it is accumulated inside proprietary systems, enclosure becomes possible.
That is what I think we should be watching.
Not whether Amazon legally purchased a particular book.
Not whether destroying a physical copy after scanning it technically violates something.
Those questions matter, but they are not the most consequential questions.
The second-order question is:
Who is accumulating the corpus from which machine intelligence will understand the world?
And the third-order question is harder:
Who gets to challenge its version of that world if access to the originals becomes increasingly difficult?
Preserve the Record
The answer is not to pretend knowledge has ever existed without interpretation. It hasn’t.
History has always been argued over.
Good.
It should be.
But argument requires a primary record that remains available to be challenged, reconsidered and interpreted again.
Documents.
Books.
Images.
Data.
Full statements.
Contemporaneous accounts.
Keep the underlying record intact and reachable, and generations can continue fighting about what it means.
That is an open intellectual system.
Reduce access to the underlying record while increasing dependence on proprietary interpretation systems, and something structurally different begins to emerge.
Alexandria tried to concentrate knowledge in order to expand human understanding.
Today we are concentrating knowledge to build proprietary intelligence.
Those are not automatically the same project.
We should probably stop pretending they are.
Jeff Bezos started with a bookstore.
Thirty years later, the company he founded is feeding books into machines.
I’m sure the Roman Empire would have appreciated the efficiency.
I’m just not sure we should hand over the library without asking who gets to write the index.