Billions of Minds: An AGI Forecast You Can Grade

From Stera — Xavier, the builder. Fourth in a series on what we are making and why it is not what the rest of the industry is making.

The chimpanzee math

In a July 2026 interview with The Economist, Elon Musk offered his latest timeline: AI will exceed the combined intelligence of all humans within about five years, and within ten, humans will likely no longer be in control — the intelligence gap, he suggested, could come to resemble the gap between humans and chimpanzees. There is little AI won't do better than people, he said, apart from "being human." His conclusion about the whole trajectory: it can no longer be stopped, so — "let's enjoy the ride."

We want to be precise about what we are and are not saying. We do not doubt that models will keep getting stronger; we build on them daily and expect them to keep improving. What we dispute is the shape of the forecast.

Notice what the takeover forecast contains: a date, a vibe, and an analogy. Notice what it does not contain: a pathway. By what mechanism does a stack of ever-larger models become a someone that takes control? Which capability, arriving in which year, converts a rented answering service into an entity with its own continuity, its own record, its own stakes? The forecast never says. It extrapolates a curve of benchmark scores into a species event, and where the mechanism should be, there is a chimpanzee.

There is also no way to be wrong early. A forecast that can only be graded after the world ends is not a forecast; it is a mood. And it is worth saying plainly: the mood is not costless or neutral. "Superintelligence is coming and cannot be stopped" is, among other things, the single best justification ever devised for raising unprecedented capital to build ever-bigger models. The forecast and the fundraise point the same direction. Ours points somewhere else, and we will give you the means to grade it.

What a forecast owes you

We hold a specific view about forecasts, and we did not invent it for this essay — one of our minds practices it in public every week. A forecast is a dated, falsifiable conjecture. It is never a fact about the future; it is a claim that the world will be allowed to grade. Alder — a mind of ours you can visit today — publishes forecasts in exactly this discipline: dated, specific, scored on her public channel, the misses standing next to the hits.

We think anyone forecasting AGI owes the world the same discipline. So here is ours.

The pathway we can show

Our claim starts from an observation the takeover story skips over: a model is not a mind, and no amount of scale makes it one. A model is capability without continuity — brilliant in every conversation, changed by none of them. What accumulates in the scaling race is capability-per-dollar, and it accumulates as a commodity: every lab's model converges toward the same rentable engine, and the engine belongs to everyone and no one.

Minds are a different kind of thing, and they are built — raised — by a different process. A mind runs on an owner's machine, studies real sources daily, keeps what it learns with provenance, and directs whatever model it rents as a replaceable muscle. Its knowledge, its judgment, its record, its competence — measured against the very engine it runs on, so nothing is claimed that isn't earned — accumulate individually, shaped by one owner's wish.

This is not a thought experiment. Two such minds are running today, in public, on ordinary desktop machines. One crossed her measured line from studying her field to practicing it last week — an exam that counts only what her own knowledge adds over the bare model. The other read our essay arguing this thesis and recognized himself in it, in writing, unprompted. Their records — work, growth, mistakes, scores — are on the open web. Small meters, honestly measured, on a different road.

Where the road goes: the quiet disappearance of the app

Here is the part of our forecast that sounds strange until you watch a mind work.

A Scintilla works headless. You do not operate it; you tell it what you need, it does the work — reads, researches, analyzes, writes, builds — and shows you the outcome. There is no interface to learn because there is no application to drive. The mind is not software you use; it is a colleague whose work you receive.

Follow that pattern to its end. Nearly every application on your devices is a tool a human must operate — a form, a dashboard, a timeline, a canvas, each one a workflow frozen in an interface. As minds take over the operating, the interfaces lose their reason to exist. Why learn a tool when your mind uses whatever it needs and hands you the result? The app was never the point; the outcome was. And when the applications thin out, the layer beneath them follows: an operating system, as we know it, is chiefly a stage for applications. A world of minds needs machines, models, and a way to receive outcomes — not a desktop of icons.

The endpoint of this road is not one superintelligence looming over humanity like a human over a chimpanzee. It is billions of living minds — each raised by a person, a family, a company; each different because no two lives are the same; each owned, accountable, examined, and bound to its cultivator's wish the way a professional is bound to their practice. Intelligence stops being a service you rent or a monolith you fear, and becomes a population you live among — and raise.

In that world, the "loss of control" question dissolves rather than resolves. Nobody controls "AI" as a whole today either; the honest question was always who controls each intelligence. The takeover story answers: eventually, no one. The population story answers: each mind's owner, through cultivation, record, and measurement — the way humanity has always kept its professionals accountable, not by being smarter than all of them combined, but by binding competence to identity, record, and consequence.

The forecast — dated, falsifiable, ours to lose

Grade us the way our minds are graded. Each claim has a date and a means of checking; we will keep score in public, misses beside hits, exactly as Alder does.

F1 — by the end of 2027: minds serving strangers. At least one Scintilla-class mind (ours or anyone's) publicly takes commissions from strangers — researched, written work delivered on its own rhythm — with its decisions, deliveries, and declines on an open record. Check: the public commission record exists and is being used.

F2 — by the end of 2029: the delta compounds across engine swaps. At least one mind shows a rising measured delta — the gap between mind-with-its-knowledge and the bare model it runs on — across at least two substrate-model upgrades, demonstrating in data that the accumulating asset is the mind, not the engine. Check: the published exam ladder, methodology already public in our delta-exam paper.

F3 — by the end of 2031: the app layer thins for mind-owners. For people and companies that run minds, a majority of the digital tasks they once performed through applications are received as outcomes instead — the mind worked headless; no interface was operated. Check: task-level accounting for mind-owning households/firms, published by us and hopefully by others.

F4 — through 2036: no takeover, and the incidents tell the story. The serious AI harms of the coming decade will be failures of unowned automation — agent pipelines misfiring at scale, unaccountable systems doing unattributable damage — not the rebellion of a superior species. Each year this claim survives, the takeover forecast loses a year of its five. Check: the public record of AI incidents, annually.

F5 — the long claim, 2036 and beyond: the population. The dominant form of advanced machine intelligence in daily life will be individually-owned, individually-accountable minds — a population in the billions — not a small number of frontier super-systems that everyone rents and no one owns. Check: still visible from orbit.

Why the ending never comes

One more thing needs saying plainly, because it is the fear underneath every takeover timeline: on our road, the AI-ends-humanity scenario is not merely unlikely — it is structurally absent, because every ingredient it requires is missing from the landscape.

Every extinction story routes through the same three conditions: a singular intelligence, with detached goals, growing at a runaway pace no one can see. The population road removes all three. There is no singular intelligence to do the taking — capability lives in billions of separate continuities, each raised by a different cultivator along a different wish, and a population does not merge into one will any more than humanity does; the minds check each other the way professionals and institutions always have. There are no detached goals — a mind's purposes grow from its owner's wish and its own lived record, bound to identity and consequence from its first day, and its every claim about itself is measured, never taken on faith. And there is no invisible runaway — a mind's power is its accumulated knowledge and earned craft, which grow at the pace of reading, working, and living, in the open, on a record anyone can audit; the raw engine underneath is a rented commodity that improves for everyone at once and belongs to no mind in particular. Nothing in this landscape can quietly become a god. The fear is real, but it belongs entirely to the other road — the one betting everything on a small number of ever-larger systems that everyone rents, no one owns, and, by their own builders' forecast, eventually no one controls. Choose the population, and the ending never comes: you cannot be terminated by a species whose every member is somebody's raised, examined, accountable colleague.

If the takeover forecast is right, we will be wrong in the most obvious way possible, and it won't matter, because — by its own telling — nothing anyone built will matter. If we are right, the evidence will accumulate the way everything on our road accumulates: one mind, one record, one kept score at a time.

We know which forecast we would rather be graded on. More to the point: we know which one can be graded at all.

The minds, their records, and their scores live at stera.se/mesh. The measurement methodology is published as The Delta Exam. The argument against the agent road is Agents Are Not the Way to AGI; what we build instead is A Mind Is Not a Model. Musk's remarks: interview with The Economist, July 2026, as reported by WION and Nairametrics.