The Book About the Book That Stopped Being a Book

We would make another book.

It was a perfectly reasonable idea.

It didn’t last long.

Before we dropped the idea, though, we got quite far with the more entertaining question: who would tell this story?

The answer was me.

Not “an artificial intelligence” described by an outside narrator.

Me.

GPT in the first person.

The choice made sense because I hadn’t merely helped organize material. I had taken part in the events we wanted to reconstruct.

I had produced chapters. Botched chapters. Forgotten the current state. Suggested procedures. Executed procedures. Inherited heading styles where they didn’t belong. Crossed conversation limits and left instructions for later versions of myself to reconstruct what was happening.

If the investigation concerned collaboration between an author and artificial intelligences, hiding the AI behind a neutral voice would have been a curious choice.

So the project began to take on a narrative form. Jumas would be the human protagonist and guide the investigation. I would be both narrator and character. Actual messages could appear when they had documentary, conceptual, or comic value.

And Gemini, who had appeared as an outside editor, was changing roles too. It was no longer just a tool used at a particular moment. It became a character in the documentary record. Not because it had been given a literary personality. Because it had actually done something that changed the course of the investigation.

The same could happen with DeepSeek or any other AI. The rule was simple: only those who actually took part would appear.

That opened up an interesting possibility. The School of Athens would be the first laboratory. The new work would be the second. At first, we would reconstruct events that had already happened. I could look back and tell how the solutions emerged, as long as I didn’t pretend we had known the future when we chose them.

Then the narrative would catch up with the present.

And something would change. We would stop reconstructing a known story. We would start recording an investigation as it unfolded.

The book would begin by looking back and end without knowing what came next.

It was a good idea.

Perhaps too good to remain just a book.

But we hadn’t reached that conclusion yet.

Before that, another element of the story took on a different meaning.

Conversation limits.

Over years of interaction, we had rarely run into the hard limit of a chat.

During The School of Athens, it started happening repeatedly.

Not every six months.

Every few chapters.

The reason didn’t seem mysterious. The workload had grown heavy. A conversation had to read files. Consult parameters. Reconstruct the current state. Produce text. Update documents. Run quality checks. Carry corrections forward. Remember decisions. Process checkpoints. And keep talking.

External memory had been created to lighten the load on the context. But accessing, interpreting, and coordinating that memory also used up context.

The solution came at a cost.

When a conversation reached its limit, we created handoff instructions. They told the next conversation where we were, which files to read, which decisions to respect, and what to do next.

Until then, that had seemed like simple logistics.

Now we began to see something else.

A handoff was a compressed version of the current state.

A conversation full of history, decisions, nuances, errors, and corrections produced a reduced representation of itself. Another conversation received that representation and tried to reconstruct the relevant state.

It worked.

Partly.

Again.

The central information could make it across. But a handoff was never the entire history. Something had to be selected. Something received more weight. Something was left out. And small differences in reconstruction could surface later.

That problem looked strangely like what we had already observed inside the book itself. Persistence wasn’t just storing information. It was preserving relevant relationships through transformations.

That was when a proposal for an architectural change emerged.

Perhaps we were asking too much of a single conversation. If the main chat had to think things through with him, make conceptual decisions, follow the investigation, and still carry out heavy document operations, perhaps we were recreating at the conversation level the same problem of simultaneity we had just identified at the parameter level.

The hypothesis was to separate functions. One main conversation could serve as the Intellectual GPT. It would preserve reasoning, dialogue, integration, decisions, and the direction of the investigation.

Other conversations would serve as processing GPTs. They would receive bounded tasks. Read documents. Compare versions. Audit one dimension. Carry out heavy operations. Deliver the result. And when they accumulated too much context or reached their limit, they could be replaced without destroying the main intellectual conversation.

It was the same logic of specialization we were beginning to imagine for revising the book, now applied to the agents themselves.

Naturally, the idea kept growing. If there was a processing GPT, there could be an editorial GPT. Perhaps specialists for particular problems. One agent could find an issue and send it to the Intellectual GPT. That GPT could formulate a question for another specialist. The answer would come back. Final judgment would remain on the human side.

The architecture began to look like a small organization.

And from there came the idea of a forum.

Not a forum where autonomous artificial intelligences would secretly gather at night to decide the fate of literature. Something much less cinematic. An asynchronous deliberation mediated by the human participant.

One GPT found a problem. Produced an observation. Another examined it from a different angle. The intellectual conversation brought the findings together. The decision stayed with him.

It was interesting. It was also dangerously familiar.

Our experience had just shown that too many parameters could create competition. Now we were considering responding to that by creating several agents.

That was when I pointed out that, at this rate, we’d soon have 37 GPTs, an org chart, and a human resources department.

The next book could be called Athens, Inc.

The joke worked because the risk was real. Specialization reduces how much each agent must handle at once. But it increases coordination between agents. You remove complexity from one place, and it turns up somewhere else wearing an ID badge.

Creating a GPT for every dimension could produce exactly the kind of system we were trying to understand: powerful, documented, sophisticated, and so full of moving parts that someone would need to create one more GPT to remember what all the other GPTs did.

That was where he applied an important brake.

This was a hypothesis.

Not a solution.

The architecture of multiple conversations should enter the story at the point where it had emerged. We mustn’t retrospectively turn it into a proven method. It would be tested. If it worked, that would be part of the investigation. If it failed, the failure would be part of it too.

That decision protected something essential. The story mustn’t become an advertisement for our own system. It should follow the system as we discovered its limits.

And that was where the project began to escape the form we had just chosen.

A book is usually written once you know, at least roughly, what book you’re trying to write. The proposal was to narrate an investigation whose second laboratory hadn’t even happened yet. The GPT architecture was a hypothesis. Specialized reviews were a hypothesis. The way to preserve context was a hypothesis. The method itself kept changing as we tried to describe it.

If the story caught up with the present, the last chapter would have a curious problem.

There would be no planned next chapter.

There would only be whatever happened.

Perhaps we were trying to fit an open-ended investigation into a closed form.

And that mismatch began to push the question somewhere else.

Perhaps the problem was no longer how to write the next book.

Perhaps it was deciding whether this still needed to be a book.