Every forecaster eventually meets the date they set. Most arrange never to be in the room when it arrives. This is me, in the room, on the date — and the interesting question is not whether I was right, but which half of the prediction was wrong, and why the wrong half is the reassuring one.

I. The Man Who Set a Date

On 29 November 2025 I wrote, under my own name, that we had roughly two years left. The recursive self-improvement loop, I said, would close at superhuman level – I dated it to mid-to-late 2026 on the then-current extrapolation – and when it did, progress would run from impressive to incomprehensible in weeks. I called that moment, in as many words, the point at which humanity loses the steering wheel. I gave a band of eighteen to thirty months and I did not hedge it into meaninglessness.

Today is inside that window. So begin where a forecaster is least comfortable beginning: on his own overdue claim, before anyone else marks it for him. A prediction that costs the forecaster nothing when it fails is not a forecast. It is a mood with a date attached, and the date is decoration. I would rather the date cost me something, so here is the invoice.

The loop has not closed. Nothing I would recognise as a superhuman, autonomous self-improvement engine exists in August 2026. The steering wheel is still, embarrassingly, in human hands. Read that as a confession and you have read it wrong.

A forecast that costs the forecaster nothing when it fails is not a forecast; it is a mood.

II. What Was Claimed, and What Arrived

Be precise about what was claimed, because the value of a marked prediction is entirely in its precision. Two claims travelled together in that essay and should not have. The strong claim: an autonomous loop, machines improving machines faster than the human scientific community, closing at superhuman level inside twelve months. The weak claim: that AI systems were beginning to accelerate their own development. I wrote them as one sentence. They are not one sentence, and the year has pulled them apart.

The strong claim has not survived contact with the calendar. A July 2026 survey of the field by Chen, Wang and Qu draws the distinction I should have drawn myself: bounded self-refinement is, in their phrase, already industrial practice, while open-ended recursive self-improvement ‘remains bounded by grounding requirements, collapse dynamics, and compute constraints’. Humans still set the research direction. The autonomous loop is not late by a fortnight; it is structurally not here, held back by constraints that a calendar does not dissolve. METR’s frontier time-horizon measurement, as of early 2026, puts no model at researcher-level autonomous AI research. The measuring instruments do not show the thing my sentence asserted.

The weak claim, meanwhile, arrived on time and then some. A frontier lab has stated that one of its own coding models was, in effect, instrumental in building its successor; release cycles that ran to months now run to weeks. That is real, and it is not nothing. But it is assisted engineering, not autonomous recursion, and the gap between the two is the whole of my error. I compressed a real, bounded, human-directed acceleration into an imminent, unbounded, self-directed one, because the second sentence was more frightening and I was writing to frighten.

The capability curve kept its appointment. The catastrophe missed its own.

III. The Distinction the Countdown Missed

Here is the thing the countdown obscured, and it is worth more than the countdown ever was. The two curves I fused are not the same curve. Capability — what a system can do — advanced through the window roughly on the schedule the optimists drew. Control — whether anything can reliably stop the system from doing a thing at the moment it tries — did not fail. We did not lose the wheel. And the reason we did not is not that capability stalled. It is that loss of control was never a pure function of capability in the first place.

Apply the test this publication applies to everything. Can a mechanism refuse an action at execution time, deterministically, in bounded time, independently of the model whose behaviour it governs? That question does not get harder in lockstep with capability; it gets harder in lockstep with autonomy of action, which is a different axis. A model can become enormously more capable while the number of things it is permitted to do without a gate stays exactly where an engineer set it. The twelve months just past are the cleanest natural experiment we could have asked for: capability climbed, and the harm the countdown promised did not materialise, because the binding constraint on an agent is not how clever it is but what it is allowed to execute.

I predicted the wrong variable would bind. I watched the capability curve because it is the one with a graph. The variable that actually governs sat in the boring place — in what the deployment permits — and it held, not because anyone had built the enforcement layer properly, but because agents in 2026 are still mostly asking permission by construction. That is not safety. That is a reprieve.

We did not lose the wheel because the wheel was never bolted to the engine; it was bolted to the throttle.

IV. Feel the AGI, Two Weeks Earlier

There is an inconvenient document to deal with before I go further, and it is also mine. Two weeks before I set the countdown, I published an essay arguing that fixed AGI timelines behave like a conspiracy theory — that a precise date for a civilisational threshold is a tell that someone is selling something. Then I turned around and sold you a precise date for a civilisational threshold. A reader is entitled to ask which of the two men writing under my name to believe.

The honest answer is that they were answering different questions and I let them blur. Capability prophecy — the specific month the loop closes — is exactly the conspiracy-shaped object my earlier essay warned about, and my countdown was an instance of the error I had just named. That is a genuine contradiction and I am not going to smooth it. But the enforcement argument does not depend on the timeline being right. Whether the dangerous capability arrives in 2026 or 2032 — and the field’s own current extrapolations point closer to the latter — the question of whether anything can refuse an action at runtime is the same question, with the same answer, on either date. The prophecy was advisory and it was wrong. The control argument is binding and the date does not touch it.

Confuse the thing you can measure with the thing that governs, and you will forecast the wrong apocalypse on time.

V. What I Got Wrong, With the Arithmetic

Let me state the reduction plainly, because a concession that is not stated as a reduction is just an apology, and I am not apologising for the argument, only correcting its shape.

I claimed: autonomous superhuman self-improvement, mid-to-late 2026, loss of human control. I now claim: bounded, human-directed self-assistance is real and accelerating; autonomous self-improvement is not here and the primary evidence says it is constrained by more than compute; and the loss-of-control scenario I attached to the timeline was attached to the wrong axis. What survives is smaller and, I think, truer: the window in which we can install runtime control cheaply is finite, and we are spending it. What does not survive is the date — and, as the case against this essay will force me to concede, two things more: that control held this year for reasons that had nothing to do with any governance I or anyone else deployed, and that a window without a date owes a falsifier I have not yet written. Hold me to both.

I should tell you where my interest sits, in the body, at the point it becomes convenient, because that is the rule this publication holds others to. I build runtime-control infrastructure. An argument whose conclusion is the binding variable is control, not capability is an argument for the thing I sell, and you should discount it accordingly. Then weigh the discount against the evidence, which comes from METR and from a survey by authors with no stake in my company, and against the fact that the person the essay embarrasses is me. I have made the case less convenient for myself than for my competitors, which is the only credential a marked prediction carries.

I was wrong about the date. I would rather be wrong about the date than silent about the variable.

VI. The Case Against This Essay

The danger of a mea culpa is that it purchases trust cheaply and then spends it on the conclusion the author wanted all along. Four objections, and the second is the one that lands.

The salvage objection. The essay concedes the date and then quietly keeps the frightening frame, re-describing a plain miss as a half-hit by splitting one prediction into a wrong strong claim and a correct weak one — a split I did not make in November and am making now precisely because it rescues me. Answer, partial. The split is real and defensible in the sources, but the objection lands on motive: I would not be drawing this distinction so carefully if it embarrassed me further. Grant the reader the right to discount the tidiness.

The incidental-reprieve objection. The essay’s own §III concedes that control held not because anyone built the enforcement layer, but because agents in 2026 still ask permission by construction. If so, the variable held for a reason that has nothing to do with the infrastructure I sell — the reprieve was a property of immature deployment, not of governance — and the essay cannot claim its own solution was vindicated by the window it just described. Conceded. This is correct and it is the real cost. The most I can claim is that the problem is genuine and not yet urgent; I cannot claim my answer to it was tested, because it was not deployed. That is a materially weaker commercial position than the essay otherwise implies, and I have to sit in it.

The unfalsifiable-successor objection. Having lost a dated, testable claim, the essay replaces it with an undated one — a window that ‘closes quietly, one permitted action at a time’ — which cannot be marked to market by anyone, ever. That is not wisdom; it is a mood with the date filed off. Answer, partial. Fair, and the remedy is a duty, not a rebuttal: the successor claim owes its own falsifier, and I owe it in a follow-up — a specific, measurable threshold of agentic autonomy-of-action past which the cheap-install window is demonstrably shut.

The interest objection. The conclusion favours the author’s business. Answer. Declared in §V, and discounted there; the evidence is third-party and the embarrassment is first-party.

A confession that leaves the confessor better positioned than the truth is just marketing in a hair shirt.

VII. Coda: The Clock

Return to the clock. It has run most of the way down and the alarm I set did not sound, and the temptation now is the comfortable ending — I was wrong, it is fine, panic was unwarranted, carry on. Do not take it. The clock was measuring the wrong quantity, and a broken clock that stopped at the wrong hour tells you nothing about whether the building is on fire; it tells you your instrument was pointed at the wrong wall. The window I actually care about — the years in which runtime control can still be installed into agentic systems cheaply, before autonomy of action outruns it — is real, is finite, and does not announce itself with a countdown. It closes quietly, one permitted action at a time, in a room where no alarm is set. I set my alarm to the loud thing. The quiet thing is the one that governs.

The alarm I set was pointed at the wrong wall. The fire it missed is still, patiently, a fire.

---

Sources

• Soon, L., We Have Roughly 24 Months Left of Human Control Over the Future, Genesis Human Experience, 29 November 2025. src-soon-24months-2025
• Chen, M., Wang, L. & Qu, B., Recursive Self-Improvement in AI: From Bounded Self-Refinement to Autonomous Research Loops, arXiv 2607.07663, 8 July 2026. src-rsi-survey-2026
• METR, Time Horizon 1.1, 29 January 2026. src-metr-th11-2026

Provenance note: a widely repeated ‘>10x speed-up is unlikely’ gloss on METR’s model evaluation, a projected ~2032 date for near-total AI-R&D automation, and the specific wording of the frontier lab’s self-improvement admission are all reported at second hand and are held out of the argument above until the primary pages are located. They shape the reading; they do not support a sentence.