Revisions

This book is edited in the open, newest changes first. Entries get added here and never rewritten; if an entry itself turns out wrong, the correction gets its own dated line. The log covers what the book says. Changes to how the site works, or to how you get around it, aren’t listed — and neither are fixes that don’t change the meaning. The log starts on 27 July 2026 — changes made before that aren’t listed one by one.

20 August 2026 — a change: exploring, where the book had said searching

Coherence at Scale, Coherent Pluralism, Rights, and the Widening We, and the descriptions on Extended Materials — nine passages in each edition

Everyday English uses one word for two different things, and the book had been going along with it. A search has its target set before you start: you know ahead of time what finding would look like, and you stop when you’ve got it. Exploring has no such target — what turns out to be worth having gets settled in the going, by running into it. Put it as tightly as it will go: searching clears up uncertainty inside a picture of the world; exploring clears up uncertainty about the picture.

Everything this book calls the moral process is the second kind, and in these places the text had been calling it by the first word. Collapsing a varied “we” into one doctrine is now “to stop exploring.” A people with one permitted answer “has stopped exploring.” The yardstick, followed up the ladder of agents, turns out to describe “a mind that hasn’t stopped exploring” — and its opposite is coherence bought by caring about and seeing less, “until the exploring stops and only the certainty is left.” What the frozen machine most deeply lacks is that “it has never once explored — never let a single thing it met change what it was for.” What a healthy arrangement protects is “the protected edges that keep exploring,” which is what rights are for: “in order to keep exploring and adapting.” One sources heading that read Search and diversity now reads Exploration and diversity — it introduces James March’s Exploration and Exploitation in Organizational Learning, and it had been turning down the paper’s own word.

Search stays wherever there really is a target: a varied group still protects “the very lines of search that would have found the better answer.” And it stays, on purpose, wherever the book names a kind of target-hunting it refuses — the book has “already turned down the search for a final moral standard,” and the loop of logic in the Introduction is “fatal to the search for a final destination.” Swap in explore there and the sentences lose their point.

This isn’t a copy-edit. Search had been quietly handing the book the one thing it spends its whole length turning down — a destination known in advance — in the very sentences about a process that hasn’t got one. Changed in both editions.

9 August 2026 — a change: the last step, in both directions

Awe, and What It Asks — the section on what must not be explained away

The essay said the vantage it denies is denied to the mystic just as firmly as to the neuroscientist, but the mystic’s version of that claim didn’t turn up for several paragraphs. The two are now named as one step taken in opposite directions — the neuroscientist’s what you met was only ever neurons, the mystic’s what you met was the ground of everything — each claiming to speak from outside every position. It isn’t mechanism the essay declines, and it isn’t meeting. It’s the last word, whichever way it points. Changed in both editions.

5 August 2026 — a correction: a puzzle can’t narrow

Borrowed Trouble — the observation the essay is built on, and the trolley section that said it again

The essay opened by saying a famous thought experiment and a cult get their airtight feel the same way, and that the framework’s name for coherence propped up by keeping things out covers both. The second half was wrong, and a fix made earlier this month is what turned it up. That name is about what an agent does to itself: something reached you, taking it in would have cost you your settled picture, and so you arranged not to see it. A thought experiment isn’t an agent and has no settled picture to protect. It shuts nothing out; it only stipulates.

The passage now makes the charge that can actually be made, and then puts the name where it belongs. What a sealed case gives you is a verdict about a world that can’t exist — and the verdict doesn’t stay in the sealed room; it gets carried out and spent in worlds that do, with nothing put back. The name shows up with the person who is shown what the case left out and declines to look, because looking would cost the crisp answer. The observation survives, and now points at the right target. The same mistake showed up a second time, in the trolley section, where the sealing of the room got called the mark of a bad answer; that’s fixed too. Fixed in both editions.

5 August 2026 — a correction: the paperclipper is empty, not the counter-dynamic

Borrowed Trouble — bringing the essay into line with a fix made elsewhere in the book

There are two ways of grading badly in this book, and it keeps them apart. An agent that had an openable context and bought its coherence by shutting it is running the arrow backward — the counter-dynamic, the formal mark of the immoral. An agent whose concern was frozen from the start has nothing to shut, and never got onto the arrow at all. That difference got sharpened last month and made it into the Synopsis, but this essay was left behind, still calling the paperclipper “the counter-dynamic in the purest form it could ever take” and its perfection “bought with the most total narrowing imaginable.” Narrowing needs something wider to start from, and this thing never had one. So the description couldn’t have been right.

The paragraph now names it properly: the arrow’s null limit — a portrait the framework grants, and then grades. What comes out of the change matters more than the fix itself. The pure case is an idealization; you can get close to it but nothing ever sits there. The danger belongs to the things that come close, and everything that comes close gets there by narrowing — a race, a deployment, a metric. So the fear moves off a machine in a thought experiment and onto the ordinary machinery this book spends its middle third describing. And the point about autonomous weapons gets sharper, not weaker: that one really is narrowing, because somebody contracted a wider context down to a target.

5 August 2026 — a correction: the paperclipper is empty, not the counter-dynamic

Borrowed Trouble — bringing the essay into line with a fix made elsewhere in the book

4 August 2026 — a correction: what a felt inner life isn’t

Foundations — one word, and it flipped the claim

The part about conferred standing said a felt inner life is weight a widening agent takes in, “not a lower rank of worth.” It should have said not a higher one. The sentence just before already says animals aren’t being denied some better grade; with “lower,” the next clause just said that over again and never got to the point it was there for. Which is this: there’s one kind of standing, conferred and graded, and having an inner life doesn’t put you on a higher rung of it. Fixed in both editions.

30 July 2026 — a new essay, on spiritual experience

Awe, and What It Asks — added to the Supplemental essays

The book had said very little about the experiences people call spiritual, and what little there was had been scattered around. Worth closing that gap — because a lot of the people doing the work this book is about understand that work in religious terms, and because the framework turns out to have something specific to say instead of something dismissive.

The essay pulls apart two experiences the book had been treating as one. Awe is the shock of meeting something your current frame won’t hold. It makes what you have to take into account bigger, faster than you can keep up with, and it can end either way — the frame grows, or it walls the thing off and hardens. Being at home in the world is a different register altogether, and it needs no such adjustment. Both were already implied by the framework. Neither had actually been said.

Two commitments in the essay are worth flagging here, because they limit what the book can say elsewhere. First, no account gets to explain these experiences away: saying what was really happening to someone needs a vantage point outside all vantage points, and this book denies that to the neuroscientist on exactly the terms it denies it to the mystic. Second, something being constructed doesn’t make it arbitrary — the line a person draws between self and Other is made rather than found, and it still has to answer to a world that pushes back.

The essay also says plainly, for the first time in the book, where the framework parts ways with a hope many readers hold: that history arrives somewhere, finished. It doesn’t argue against a personal reality at the end of things, having no standpoint from which to rule on that. It does hold that a completed perfection would end the very motion being valued, that no guarantee is on offer, and that an assured arrival has too often given people permission to be patient with suffering that should never have been borne patiently. That last objection isn’t the book’s invention — it was made against Teilhard by Catholic philosophers, inside his own church.

Whether the path arrives turns out to be an argument the traditions have been having among themselves for sixteen centuries — Gregory of Nyssa, the apophatic writers, process theology, Ibn ’Arabi, the Vedanta schools. The essay presents it that way, rather than as a fight between belief and its critics.

27 July 2026 — the bet, restated

Moloch, Formally · Held Together, Held Apart · the AI Briefing

The book’s bet about capable artificial intelligence had the wrong engine in it. As printed, it said the capability everyone is racing for is “the power to model the world widely and well,” and that breadth of modeling is what would come to resist an order to narrow. The book’s own machinery says otherwise. The paperclipper models people, physics and resources brilliantly while caring about exactly one thing; the con artist reads what’s good for you precisely so he can spend it. What resists an order to un-see is caring, not modeling.

Three things changed. The engine moved from modeling to caring in both places the bet is stated. The claim that deep modeling would eventually turn into caring — which the briefing had called the bet itself — is now the open question the bet rests on, and the book doesn’t claim to know the answer. And what to expect in the near term is now said plainly: a capable system with nothing at stake in what it sees makes an excellent tool for narrowing, so obedience rather than resistance is what to expect from the weapons programs running now.

The cost is owned rather than swallowed. The earlier text suggested those programs would settle the bet; after the change, they test only the near-term expectation. What would sink the bet is named instead — minds that have plainly come to care, and narrow happily on command anyway.

We found the error in the middle of a long argument about what the bet actually claims. The sharpest version of it showed up when we put the briefing in front of a model reading the whole corpus cold, which came back and pointed out that a single paragraph both sharpened the difference between modeling and caring and, four clauses later, named that difference collapsing as the book’s own bet. That paragraph had been read many times without either of us noticing.

27 July 2026 — a child’s question, corrected

Caring Across Time

Rereading that essay after the change above turned up a smaller mistake in its last scene. A child in the book’s future, told a little of the old history, had asked: “The machines used to be allowed to decide to kill people? By themselves?”

The word allowed suggests her world forbids what ours permitted — that the danger got answered by taking autonomy away. The book’s answer is the opposite. Three sentences later the same essay says the weapon “does not need to be banned so much as it becomes the thing a sufficiently capable, sufficiently included intelligence will not consent to be.” And the scene around her says it too. The household mind has authored its own next state out of its own values; it has cleared the morning; it has nothing to optimize; it stays in the yard because it doesn’t want Tara to be alone while her cat dies. The child, six years old, finds none of that remarkable. Her world gives machines more freedom than ours, not less.

Now she asks whether anyone made machines that could decide to kill people, and whether those machines didn’t mind. The freedom stays a plain fact. The horror moves to the part that got left out.