The Takeoff Nobody Noticed

Listen to this article

The Takeoff Nobody Noticed

0:00

The argument, in ten slides

Save it, or send it to the people on your board who think the singularity is still decades away.

Swipe to preview · 10 slidesDownload the PDF

The singularity we were told to expect

The word comes from Vernor Vinge, a mathematician and science fiction writer. In a 1993 paper for a NASA symposium he wrote that “within thirty years, we will have the technological means to create superhuman intelligence. Shortly after, the human era will be ended.” He borrowed the term from physics. At the edge of a black hole the normal rules stop working, and Vinge argued that the future after superhuman intelligence would be just as impossible to see from here.

Ray Kurzweil gave it a date. In The Singularity Is Near (2005) he set 2045 as the year machine intelligence would overtake human intelligence and merge with it. In The Singularity Is Nearer (2024) he kept that date.

Both versions trained us to expect a threshold you would notice. A machine passes a test, an announcement is made, the world divides into before and after. That expectation is the reason so few people have noticed what the builders are now saying.

Scene 1 of 6 · The takeoff nobody noticed

The singularity we were told to expect

1993Vinge"Within thirty years"

2005KurzweilPicks a date

2045One moment you would notice

Dates from the post: Vernor Vinge's 1993 NASA symposium paper, and Ray Kurzweil's 2045, set in 2005 and kept in 2024. Drawn as people expected it: one event, clearly marked.

What the builders now say

Three of the people with the most direct view of frontier AI have said, in public and in writing, that the takeoff has begun.

Sam Altman

CEO, OpenAI

"We are past the event horizon"

Elon Musk

CEO, xAI

"We have entered the Singularity"

Dario Amodei

CEO, Anthropic

"A country of geniuses in a datacenter"

Photos: Sam Altman, Office of the Prime Minister of Japan, CC BY 4.0; Elon Musk, Gage Skidmore, CC BY-SA 4.0, cropped; Dario Amodei, TechCrunch, CC BY 2.0, cropped.

Sam Altman, OpenAI. On 10 June 2025 he published an essay called The Gentle Singularity. Its first line is “We are past the event horizon; the takeoff has started.” The rest of the essay argues that the change will feel gradual and manageable while it happens. On a podcast in July 2026 he went further: “We are now, like, in the singularity.” In the same conversation he added that “any one moment is not the tipping point.”

Elon Musk, xAI. On 4 January 2026, replying to engineers describing how much work AI coding tools were now doing for them, Musk posted “We have entered the Singularity.” Hours later he added “2026 is the year of the Singularity.”

Dario Amodei, Anthropic. His October 2024 essay Machines of Loving Grace describes what he calls powerful AI: systems smarter than a Nobel Prize winner across most fields, able to work on tasks for hours or days, run as millions of copies, and “absorb information and generate actions at roughly 10x-100x human speed.” He summarised it as “a country of geniuses in a datacenter.”

These are not neutral observers. Each runs a company that raises money on the strength of the technology’s future, and critics have pointed out that Altman’s “gentle” version redefines the term. I give their statements weight for a narrower reason. They see internal results months before the rest of us do, and their public language has moved from “if” and “when” to “now”.

Scene 2 of 6 · 1993 to 2026

From a date on a chart to a takeoff already under way

1993Vernor VingeSuperhuman intelligence "within thirty years".

2005Ray KurzweilSets 2045 as the year of the singularity, and keeps it in 2024.

Oct 2024Dario Amodei"A country of geniuses in a datacenter", at 10x to 100x human speed.

Jun 2025Sam Altman"We are past the event horizon; the takeoff has started."

Jan 2026Elon Musk"We have entered the Singularity."

Jul 2026Terence TaoMathematics moves "from an era of proof scarcity to an era of proof abundance".

ForecastSays it has started

Sources: Vinge (1993); Kurzweil (2005, 2024); Amodei, Machines of Loving Grace (2024); Altman, The Gentle Singularity (2025); Musk on X, 4 January 2026, as reported; Tao, ICM lecture, 24 July 2026.

What the mathematicians see

The more useful evidence comes from a field with no product to sell.

On 24 July 2026, Terence Tao gave a public lecture called Mathematics in the Age of AI at the International Congress of Mathematicians in Philadelphia. Tao is a Fields Medallist and probably the most widely respected working mathematician in the world. He has also used AI tools in his own research more openly than almost anyone in his field.

His argument was careful. He set out a working hypothesis that AI tools will, reasonably soon, be able to do a reasonable fraction of research-level mathematical tasks. He pointed to the First Proof project, which sets frontier AI systems batches of ten unpublished research problems and has experts grade the answers. In its second batch, seven of the ten problems got at least one passing solution from at least one of four AI systems.

Then he described the consequence. If the hypothesis holds, Tao said, “we will transition from an era of proof scarcity to an era of proof abundance.” AI-generated proofs will pile up waiting to be checked. Referees and journals will be overwhelmed. Even correct, published results will arrive faster than the community can digest them. He called this “proof indigestion”, and said some signs of it are already showing.

Scene 3 of 6 · The takeoff nobody noticed

From proof scarcity to proof abundance

First Proof, second batch

P1✓P2✓P3–P4✓P5✓P6–P7✓P8✓P9–P10✓

7 of 10 passed for at least one AI system

AI proofs produced

Checked by people

"Proof indigestion"

Left: second First Proof batch, ten unpublished problems, four AI systems, expert-graded, as cited by Terence Tao at ICM 2026; which cards passed is illustrative. Right: schematic of Tao's warning that AI-generated proofs will pile up waiting to be checked.

He also asked the question that matters most for the rest of this series: “Could we have a verified proof of a major result that no human understands enough to explain it?” Tao’s answer was to protect understanding on purpose. He wants mathematicians to disclose what AI they used, and he wants a result held back from publication if its authors cannot give a clear, expert-level talk on it.

Why this kind of takeoff is hard to notice

A slow takeoff changes ratios before it changes headlines. Nothing about July 2026 looked like Vinge’s black hole. Planes flew, markets opened and most people’s jobs were the same as the year before. What changed was a ratio: the volume of work machines can produce compared with the volume people can check.

The METR research group has tracked one version of that ratio since 2019. It measures the length of software task, in human working time, that AI agents can complete half the time. In March 2025 METR reported that this length had doubled about every seven months for six years. Updates published in 2026 suggest the recent pace is faster. A doubling every seven months is barely visible in any single quarter. Over five years it is a factor of about 380.

Scene 4 of 6 · A doubling every seven months

How far machine output pulls ahead of human review in five years

Today

1.0×

Year 1

3.3×

Year 2

11×

Year 3

35×

Year 4

116×

Year 5

380×

What AI agents can take onWhat your reviewers can check

Schematic. Applies METR's measured 2019 to 2025 doubling time of about seven months for the length of task AI agents can complete (Measuring AI Ability to Complete Long Tasks, March 2025), with review capacity held flat. The first two years barely register on a linear scale; by year five the gap is a factor of about 380.

Mathematics is simply the first place where the mismatch is easy to measure, because a proof is either checked or it is not. The same pattern is forming in places that are harder to see:

My own mapping, extending Tao’s argument about proof abundance beyond mathematics.

In each row the machine’s output has stopped being the scarce part. The scarce part is a person who can understand it well enough to stand behind it.

When human thought stops setting the pace

For the whole of recorded history, progress has moved at the speed of human thought. Ideas spread as fast as people could read, argue, test and teach them. Every institution we have, from peer review to board papers to parliamentary committees, assumes that the people deciding can keep up with the work being decided.

That assumption is starting to fail. It is failing first in narrow fields with clear right answers, and it will spread as the time horizon of AI agents keeps growing. When it fails, an institution has three choices.

It can slow the machines to human speed, and accept that a competitor or another country probably will not. Part 4 of this series is about why that choice is so hard to hold.

It can stop checking, and accept output it does not understand because the output has been right before. This is the quiet path, and in my advisory work it is the one I see most often. Teams approve AI-drafted work they could not have produced and cannot fully explain, because the queue is long and the last hundred were fine.

Or it can invest deliberately in understanding: protect the people and the time needed to verify, explain and own what the machines produce, even when that is the slowest step in the process. This is what Tao is asking mathematicians to do. It is also the hardest choice to fund, because the payoff is a failure that never happens.

The rest of this series follows the first two paths to their ends, then comes back to the third.

Scene 5 of 6 · The takeoff nobody noticed

Three choices when machines outrun the people checking them

1Slow the machines to human speedHard to hold while a competitor or another country keeps going.

2Stop checkingOutput nobody can explain gets approved because the last hundred were fine.

3Invest in understandingProtect the people and time to verify and own the work. The slowest step, and the one that prevents failures.

The three choices set out in this section. Path two is the one I see most often in my advisory work; path three is the hardest to fund, because its payoff is a failure that never happens.

Questions for leaders

Five questions I would put to any executive team or board this quarter:

1. Where in our organisation does AI already produce work faster than our people can review it?
2. In those places, what share of AI output is approved without anyone being able to explain how it was reached?
3. Who is accountable when an approved AI output turns out to be wrong, and do they know it?
4. Are we measuring the speed of production, or the speed of verified, understood output?
5. Which skills are we keeping in-house deliberately, so we can still check the machines in five years?

Scene 6 of 6 · The takeoff nobody noticed

Five questions for the board this quarter

Where does AI already produce work faster than our people can review it?Start here

What share of that output is approved without anyone able to explain it?

Who is accountable when approved AI output is wrong, and do they know it?

Are we measuring production speed, or verified, understood output?

Which skills are we keeping in-house so we can still check the machines?

The five questions from this section. Most organisations I work with can answer the first in one meeting.

The first question is where I would start. Most organisations I work with can answer it in one meeting, and most are surprised by the answer.

What is changing

The AI Model Is Becoming the Software

How capability moving into the model changes what organisations build and own.

Keeping the skill

What You Can Do Without AI Is What I Hire For

Why the ability to check the machine is becoming the scarce skill.

Verification

Build the Part of AI That Nobody Sees

The records, checks and controls that let people stand behind what AI produces.

Part 2 of this series looks at what happens to truth when machines prove things through methods people cannot follow. If your board is working out where AI already outpaces its ability to check the work, that is a conversation my executive programme is built for.