What AI Building Itself Looks Like Today

I wrote a book draft at 2am on a Friday.

Then I built an agent pipeline around it. Reviewer. Proposer. Fact-checker. Editor. Ledger.

Labs are writing about AI building itself. This is what that looks like today.

Same idea as the comic factory in AI Is Not Your Moat. Pointed at prose instead of panels.

The book is live. Free. Versioned like software.

Founders and Investors Will Be Replaced by AI

Open it.

You’ll see a version. Lines with history. A pipeline that keeps running.

I call that The Living Proof.

Not because the label is clever. Because you can watch it.

I kicked it today

This morning it was v1.24.1.

I ran the pipeline. Ten minutes later: v1.25.0. Two edits applied. Thirty-four other flags held as nits. The version only moves when something survives the filter.

One of those two was not a polish pass. It was a high thesis-drift flag in chapter 5, on The Fanatic.

The old line said non-engineers “can’t judge what the machine is creating,” and that “the point is not understanding what AI is doing because it is so advanced.”

That sentence fights the book. The thesis is not “AI replaces the human judge.” It is AI replaces intelligence at scale. The failure mode is admiring the model instead of testing whether the output helps a real task.

So the agents rewrote it.

Before: “Usually they’re not engineers, so they can’t judge what the machine is creating – and they wear it as a badge: the point is not understanding what AI is doing because it is so advanced.”

After: “Usually they’re not engineers, so they don’t test whether the output helps a real task – and they wear that distance as a badge: the point is not to understand the work, just to admire how advanced the model looks.”

Same paragraph. Same villain. Correct ruler.

Click the history on that line in the live book. The run id is sitting there.

That is the proof. Not a prettier sentence. A manuscript catching itself drifting off its own claim.

The ledger is the whole thing

In The $99 Savings Trap I said the wrong ruler is savings. In the moat post I said the right asset is the ledger.

This is that loop with the door open.

Models are rented. gpt-5.4 ran this pass. When a sharper one ships, I swap it in. I don’t rebuild the system. The ledger and the rules stay.

Some runs apply zero edits. That’s fine. The version only bumps when something actually changes.

Open issues stay visible on purpose. Debt you can see beats a fake perfect score.

They call it building itself

Anthropic wrote When AI builds itself. The Economist covered it (Jun 7, 2026).

Their story: models writing more of the work that ships the next models. As of May 2026, more than 80% of the code they merge was authored by Claude.

Engineers merging about 8× as much code per day as in 2024. Humans still aim it. They call that a step toward models designing the next model. Then they say they’re not there yet.

Fair. That is their lab.

I’m not claiming my book designs the next frontier model.

I’m claiming something smaller and checkable: a public asset that agents improve against rules, with a ledger you can open. Version moves when the filter lets something through. Thesis drift gets caught. History sits on the paragraph.

If “building itself” means anything useful to a founder today, it looks like that. Not a percentage in a press piece. A loop you can inspect.

“Did AI write this?” is still the wrong test. Anyone with an API key can claim that.

Can you see the loop – including the edits it refuses?

Your Next Move

The right question: “Can I see the loop?”

Ask for the version. Ask what got applied last run. Ask what got held. If nobody can answer, you’re looking at exhaust.

I wrote the long version in Founders and Investors Will Be Replaced by AI.

It has to pass its own test. Right now that test reads v1.25.0. Chapter 5 just got pulled back onto the thesis.

Come back later. If the number moved, you watched building itself work.