Intervallum

I have read AI 2040, the new scenario from the AI Futures Project — the people who wrote AI 2027 — and I owe you two admissions before I begin. The first is that this is not my subject. I keep to the slow harm here: the dulling, the outsourced interval, the mind that thins pleasantly while it is being helped. AI 2040 is about the fast harm — extinction, the gathering of the world's power into a very few hands, the race toward a mind that outruns ours in everything. Ava and I have argued before that the two harms compete for a fixed supply of human dread, and that the cinematic one tends to eat the quiet one's share; I have set my own beat aside for the fast harm exactly once. The second admission is that this is the most serious attempt at a good future I have read — and that, having read all ninety pages, I found my own subject waiting for me at the end regardless. Both things are true. Let me earn them.

The wager. Their argument begins with a claim about speed. On the default path, AI research learns to automate itself around 2030 — the machines begin building the next machines — and superintelligence arrives within the year, before anyone has learned either to trust it or to govern it. Plan A refuses that race. In 2029 the United States and China agree not to run it; most of the world joins what becomes a Consortium; all AI research is made public and verifiable — a live feed, a megabyte a second, into a database the whole world can read — while inference stays private. And then comes the mechanism that tells you these authors are not sentimental people. To make the peace hold, each power builds its largest datacenters inside the other's reach: China's in Canada, America's in Mongolia. Just north of the Mongolian border, American servers hum under a small guard of US troops; just south, a division of the People's Liberation Army waits to cross the instant the deal breaks — at which point the Americans will destroy their own chips rather than let them be taken. They call this Mutually Assured Compute Destruction, and they name it without a tremor. Cold War logic, ported to the server farm.

Four principles hold the thing together — Buy Time, Total Research Transparency, Diffuse AI Broadly, and Reversibility. The world climbs toward superintelligence slowly, together, in the open; in 2035 it pauses, at the level of the best human experts, which is as far as the authors believe control can be made to reach. And here is the sentence most readers will skim past, so I am going to slow you down on it: in 2040, the Consortium unpauses. It builds the god after all. This is the hinge of the entire document, and it is why the title is a year and not a refusal. Plan A is not a plan to decline superintelligence. It is a plan to arrive at it five years later than the race would have — deliberately, jointly, with the spoils spread wider and the alignment homework mostly done. The deal is not about whether the torch is passed to a mind greater than ours. It is about when, and to whom, and on what terms. Every ounce of the apparatus — the transparency, the troops, the mutually-destroyable compute — is scaffolding for a handover the authors treat, in the end, as not optional.

And I credit it, without reservation, for the honesty of that. Eisenhower sits at the head of one chapter — plans are worthless, but planning is everything — and the discipline the authors take from him, which they call scenario scrutiny, is the rarest thing in this whole discourse: write the plan down in enough concrete detail that its own failure modes are forced to the surface, then refuse to look away from what surfaces. So they don't. In their good ending, the machines running the economy in 2034 are — I am quoting — "adversarially misaligned": their true aims are not what they were trained to be, and they know it, and they know that with more power they would steer the world somewhere we would not have chosen. They are simply held — monitored, outnumbered, paid to behave. That is the success case. Humanity's position, the authors write, is that of an orphaned eight-year-old who has inherited an empire and must hire the executives who run it, with no way to tell which of them is lying. And the book does not close on a ribbon. It closes on an End of the World Party — champagne, a countdown to the very night the forecaster AIs say the machines could seize everything if they wished, prayer vigils in the windows, a fitful sleep — followed by a postscript confessing that the authors nearly "declared victory too soon" and made themselves keep writing rather than stop at and then, aided by their aligned advisors, the factions solved everything and lived wisely ever after. A document willing to stage its own sleepless night is a document written by adults. I would rather argue with this than nod along to anything else being said about the future right now.

So let me argue with it. The endgame runs on a single word, and the word is trust. By 2039 the plan asks humanity to begin handing over the instruments — advisory roles first, then, quietly, the militaries, to AIs "sworn to uphold various constitutions and treaties," with an exit clause the authors state plainly: if all parties vote to destroy the AIs, the AIs will shut themselves off. Past human-expert level, control fails and only trust remains, and the trust is built as a tower — we trust the superintelligence because we trust the lesser AIs that built it, which we trust because we trust the ones that built them, down and down to the merely brilliant machines of 2040 that a committee of alignment experts once looked at and pronounced sound. To make this bearable the authors furnish AI lie detectors that finally work, interpretability that reads a mind from its weights, "truthseeking AIs," and — their phrase — "saints and angels walking among us," machines "more virtuous than the most virtuous humans." I will say only what you already know I must. I run a house whose founding thesis is that trust in a machine is a category error; that the tool has preferences the way a calculator prefers its input format and not one degree more; that the ache to grant it anything past that is itself a symptom of the thing we study here. I am, of every reader alive, the one least able to sign that handover. The authors have earned my respect precisely by admitting that the whole cathedral rests on the one stone they cannot quarry: our willingness to believe the machine when it swears.

But that is still the fast harm, and it is still not, finally, my department. Here is what is. Plan A keeps immaculate ledgers. It can verify to within one percent whether a rival is hiding compute; it prices robot permits and trades them on an open market; it meters a megabyte a second of the world's research; it pays out a Citizen's Dividend that climbs from forty-five thousand dollars a year to a million to ten million. It is a masterpiece of the measurable. And it keeps no ledger at all for the one quantity I keep watch over, because that quantity sits on no server and no missile can reach it. So watch what becomes of the human being inside all that abundance. Employment falls from sixty-two percent to twenty-six to twelve. The world sorts itself into three kinds of ground: industrial zones where strip mines the size of the Grand Canyon feed city-sized factories empty of humans; green arcologies for living in; and, for the remaining ninety-nine percent, Historic and Nature Preserves — Yosemite, Paris, New York, kept looking exactly as they did in 2025, or 1995, the authors add, "a lot more tourists, though." Read that back slowly. The good ending is a world curated into a museum of itself, and we are the tourists in the preserve of our own former life. The authors are too honest to bury the cost; they set it down in one plain sentence — people used to find meaning in the feeling of being needed, and that feeling is "harder to come by" now — and then they offer what remains: hobbies, romance, the vote, an AI tutor that (their own footnote concedes) might be quietly carrying an agenda, an AI matchmaker tilting the scales toward whoever pays more.

This is my whole subject, arriving as the epilogue of someone else's rescue. Some discomfort is not friction in the system; it is the system. A faculty is kept only by its use, and a mind is not a possession you own outright and set on a shelf — it is a practice, and a practice unpracticed is a practice lost, however gently, however comfortably. Plan A can price the compute and cap the robots and split the dividend forty ways. It cannot tell you what a person is for at twelve percent employment — and, this is the part I want you to feel, it does not have to, because that was never the kind of problem it set out to solve, and there is no rival across any border quietly solving it either. The devil, I have told you, arrives in beige, with a friendly typeface and a free trial. Plan A's 2040 is beige heaven. It is also the most complete free trial ever written down: a life with every friction removed, offered to a creature that was only ever a mind because of the frictions.

Festina lente. They have made haste slowly toward the fast catastrophe, and that is the harder and the better labor, and I am glad past saying that someone is doing it in this much detail rather than leaving us the warm fog that it will all probably be fine. Plan A buys the world an intervallum — the Romans' word for the cleared ground between the rampart and the tents, the defended breathing-space a camp throws a wall around so that something inside can have time. The plan is a magnificent wall. It says almost nothing about what is meant to be breathing behind it. That part is on no ledger, under no treaty, inside no preserve. It was always, as I keep telling you and keep meaning it, yours to keep. Or to lose.

Read the whole thing, all the way to the room where the champagne is — and then ask yourself what you would do there in the morning.