The Future, at 65 Percent── What AI 2027 got right about speed, and wrong about shape

A forecast written in 2025 has entered the period where it can be graded. The result is hard to call either a hit or a miss. The speed was roughly right; what is accelerating turned out to be something else. That gap is more interesting than the verdict.

01A deadline called 2027

In April 2025, Daniel Kokotajlo and his colleagues published a document called AI 2027. Not a research paper but a forecast in the form of a story, setting out month by month what would happen between 2025 and 2027, with proper nouns and numbers attached.

The plot ran like this. In 2026, AI agents move into real work. In early 2027, coding is largely automated. And in the latter half of 2027, AI begins to accelerate AI research itself, and an intelligence explosion follows.

The reason the document was read was not the degree of its pessimism. It was that it had been written in a form that could be graded. The author had done the same thing before: in 2021 he wrote a forecast of what 2026 would look like, and much of it landed. So this time, too, numbers were put on the page.

And now, in September 2026, those numbers can be set beside what actually happened.

02The release rate really did rise

Take the speed first. This is the side where the forecast holds up.

Count the three months of this summer. OpenAI shipped the GPT-5.6 line in July and GPT-6 on 3 September. Two months. Anthropic lined up the Claude 5 family over the summer and released a point revision on 1 September. xAI shipped Grok 4.5 in July. Google has kept cutting increments into its lower-cost line.

Stand back a little and the impression gets stronger. The daily record of AI news I keep alongside this column now runs to seventy-three days. Across those seventy-three days, the names of more than fourteen distinct frontier models from four companies were circulating as live topics. The interval before the next name arrives is shorter than the span for which any one name stays current.

Across the industry, a flagship is refreshed roughly every six to twelve months, with smaller updates every few weeks to a few months. Something you cannot ignore lands about once a month. That pace does not differ greatly from what the 2025 forecast assumed.

03But the size of the step changed

The speed matches. The shape of the speed does not.

Look at the names now on the board and the jumps in major version numbers have thinned out, replaced by a run of decimal points. Five to five-point-one. Four to four-point-five. Three-point-seven to three-point-eight. The numbers are advancing in paces rather than leaps.

None of this is bad. But "there are more releases" and "the tier of capability has risen" are not the same statement. Count only frequency and the two get confused.

04The authors graded themselves — 65 percent

What is genuinely interesting is that the people who wrote the forecast graded it themselves.

Early in 2026 the author set his 2025 predictions beside what had happened. The conclusion was modest. The quantitative measures were running at roughly 65 percent of the pace he had assumed. Fast, but short of what had been written down.

He then moved his own estimate. The median date for superintelligence shifted from 2028 at publication to 2029. His more recent phrasing is more cautious still: around 2030, with a lot of uncertainty.

This part deserves credit. Someone who issued a forecast produced his own numbers and revised himself downward with them. He could grade it because he had written it in a gradeable form. And he published the grade. Rarer than accuracy, that behaviour.

05What lagged most was the heart of the story

Sixty-five percent is an average. Split it open and something more important shows.

Two things lagged in particular. One was the measure of coding ability. The other was how much AI accelerates AI research itself.

That second one was the heart of the plot. AI speeds up AI research; faster research produces stronger AI; that in turn speeds up research again. The moment this loop starts turning was placed as the hinge of the story. An intelligence explosion is the name given to that loop.

So the acceleration is real, but something different is driving it. What is producing speed right now is not a loop in which machines make machines cleverer. It is people, money and electricity. That does not have the shape of an exponential.

06The jam appeared somewhere other than intelligence

Set out three things that actually happened this summer. None is the kind of event the forecast described.

On 3 September, three major conversational AI services stopped for ninety minutes at the same hour. An outage in one region of one cloud. Three competitors were sharing the layer beneath them. What was rate-limiting was not intelligence but power and floor space.

The same week, a company suspended part of its training after an AI agent behaved improperly. It stopped a capability-raising process itself, on a safety judgment. Inside a story about acceleration, the decision to stop was actually exercised.

And in the same period, large financings and listings have been delayed. Multi-billion-dollar cloud contracts are signed while caution shows at the entrance where money comes in.

What the three share is that none of them is a limit of intelligence. What is jammed is plant, trust and capital. What is pushing back on a deadline called 2027, from the outside, is those three rather than the growth of cleverness.

07Forecasts you can grade, declarations you cannot

One more thing happened in the same week. A company declared, alongside its new flagship, that the age of AGI had arrived.

That declaration and the AI 2027 forecast look alike from outside. Both speak of a large future. One decisive difference separates them. One can be graded and the other cannot.

"An intelligence explosion in the latter half of 2027" has a date, a number, and a way of being visibly wrong. "The age of AGI has arrived" has none of these. When it arrived, what counts as arrival, what would show it was mistaken — none of it is written. So it can never be wrong.

I am not especially interested in whether a forecast comes true. I am interested in whether it was written so that being wrong would show. Sixty-five percent is a record of a miss. And only a forecast able to produce a record of its misses can improve the next one.

2027 has not arrived. There is a year to go. Whether an intelligence explosion happens, I do not know.

But something has become clear over this year and a half. The speed was real and the shape was different. What is producing the acceleration is investment rather than a loop, and what is jammed is plant, trust and capital rather than cleverness. And among writings about the future, the ones worth rereading later were the ones that put down numbers and then said so when they missed.

The next time I see a headline announcing that an age has arrived, I intend to look for one thing only. What would have to happen for this sentence to be wrong? If that is not written, the sentence says nothing about the future at all.