Moments like that are why I now keep a record of what the industry's leaders say.

Two voices

I built this tracker to follow the mood of the time around AGI and the singularity, the idea that machines will one day get smarter so fast that the world changes faster than anyone can keep up with (the home page). Two kinds of people shape that mood. One is the long-time thinkers, academics and writers who have argued about it for decades. The other is the people now running the companies that build it, whose dates tend to run ahead of everyone else's. Both are worth hearing. The thinkers give the long view; the leaders are the ones making it happen, and what they say is the clearest sign of what the industry expects right now.

What the ledger does

The claims page is the leaders' voice. It holds more than 300 dated predictions from fifteen people who run the companies building the most advanced AI or the chips it runs on, each in their own words, dated, and linked to where it was said. I read them for what they predict now and how that has shifted. It is not a scorecard. Where a prediction's date has passed, a small note on it says what happened, and I don't add those notes up (how the ledger works).

What they say now

Twelve of the fifteen have put a date on AGI or something beyond it. Seven of those twelve name a year before 2030 (the claims page). Among the experts the tracker follows are seven long-time thinkers who don't run AI companies. Line up each one's latest guess, and the one in the middle is 2030. (Two of them worked at OpenAI earlier in their careers, so they are not wholly outsiders, and the page says so.) So more than half of the leaders who give a date name a year earlier than the thinkers' middle one. Each leader uses their own words for what is coming: AGI, superintelligence, or, for Dario Amodei of Anthropic, what he calls "Powerful AI". They don't all mean the same thing, so the page gives each person's term beside their date.

Three have never put a date on it: Arthur Mensch of Mistral AI, Mark Zuckerberg of Meta and Satya Nadella of Microsoft. Not naming a year is not the same as saying nothing, so the page quotes their latest words instead. Mensch calls AGI "rather a direction than a point of arrival". Zuckerberg told his company's conference in September 2026 that "Delivering personal superintelligence is now within reach" (the claims page).

How the dates move

In April 2024 Elon Musk, who founded the AI company xAI (since renamed SpaceXSI), said an AI smarter than the smartest human would come "probably end of next year" (the prediction). In January 2026, at the World Economic Forum in Davos, he said it might come "by the end of this year", and "no later than next year" (the later prediction). The page keeps both, so the move shows.

Across the leaders, since 2023, five have brought their dates nearer, three have held them, and two have pushed them later (the claims page). The page draws every leader's forecasts over time, one small chart each.

What happens next

New predictions go on the page as the fifteen make them, and the picture of what the industry expects moves with them. The question I will keep asking is whether the two voices, the builders and the long-time thinkers, are drawing together or apart. You can watch it on the claims page, where every prediction links to the moment it was said.

Widened 2026-10-19: this post first covered 30 predictions by Sam Altman and Musk; I widened it the same day to all fifteen people (who the ledger covers).

Rewritten 2026-10-20: the first version of this post counted how many predictions had come true. I have since decided the ledger records what the leaders predict, not whether they were right, so the counts are gone (what the ledger is for).

Updated 2026-10-20: two more predictions by Mustafa Suleyman went on the page, so this post now says "more than 300" rather than an exact count that changes each time one is added. The exact number is always on the claims page.

I am John Fredrickson, and my site, the AGI Zeitgeist Tracker, asks whether we are in the singularity, the idea that machines will one day get smarter so fast that the world changes faster than anyone can keep up with. Its test has four parts: how fast AI is improving, whether anyone can still forecast it, how much of the work of improving AI is done by AI itself, and whether the world's economy has started growing far faster than it ever has. All four have to cross a line set in advance, and stay across for a year, before the test says yes (the test's rules).

A Headline That Could Only Say One Thing

Until 17 October 2026, the top of the home page said "Not yet." and nothing else. That is the test's answer, and it was right. But the test will give that answer every day until the day the singularity starts, so it told a visitor nothing about whether anything was moving.

What the Top of the Page Says Now

On 17 October 2026 I added a number under the answer (the change and my reasons, in the site's changelog). Each part has its own line to cross, and the number is the average of how far the parts have come towards their lines: 32 percent. The pace of AI's improvement is 43 percent of the way to its line, the share of AI research done by AI is 52 percent, and the economy is at 0, because its growth is still below its own long-run average. The forecasts part cannot be scored yet, so the average is over the other three.

Under the number, one sentence says where each part would go if current trends simply continued: AI's share of the research would cross its line in 2027, the pace somewhere between 2028 and 2048, and the economy not at all. So on current trends the test is not met (the home page).

Why the Economy Is Always Named Beside It

An average can mislead. One part racing ahead could make the number look healthy while the others stand still, and the rules I wrote for the test warn against exactly that. So the number is never shown without the part furthest behind, and today that is the economy, at 0.

That matters more than the average. The test passes only when all four parts cross their lines, so with one part not moving at all, the test has not begun to pass (how the number is worked out). The 32 percent is not a forecast. It shows which parts are moving and which are not.

The number also decides nothing. The answer is still worked out from each part against its own line, and it still says "Not yet." It will stay that way until all four parts have crossed and stayed across for a year (the numbers as they stand).

I am John Fredrickson, and I built this site, the AGI Zeitgeist Tracker, to follow that argument. "Zeitgeist" is a German word for the mood of a time: what people are saying and what they expect. The site records what two groups have said about AGI, with every quote dated and linked to where it was said. One is the academics and writers who have talked about this for decades. The other is the people who run the AI companies, who are actually building it and who tend to run well ahead of the evidence with their predictions. Both belong in the picture, and the point is to see how the talk has moved, not to score anyone.

The talk has moved a long way. Geoffrey Hinton, a computer scientist whose work underlies today's AI, said "no sooner than 2070" in 2015 and "between five and 20 years" in 2025. Gary Marcus, a psychologist and the best-known critic of today's AI, said "at least thirty or fifty years" in 2016 and "10-25 years" in 2025 (what people have said). Beside that record, the site asks a harder question, are we in the singularity yet, and sets a test that anyone can check.

Why the Test Needs Four Parts

Each of the classic descriptions of the singularity, taken alone, gives false alarms. A fast decade in any industry looks like runaway progress. A company can say its AI now does its own research, and nobody outside can check. Every new technology makes old forecasts look foolish, the internet included. So the test asks for several signs at once, and adds one that is measured outside the AI world: whether the world's economy has started growing far faster than it ever has. I set the line each part has to cross on 8 October 2026, before looking at where the numbers stood, so that I could not move the lines to suit the answer.

The four parts each come from a well-known thinker:

• The pace. From the futurist Ray Kurzweil, the idea that AI gets better faster and faster. METR, a research group that tests AI systems, gives them computer-programming jobs of known length, measured by how long a skilled person would take, and records the longest job the best AI can finish by itself. Those jobs keep getting longer. In 2019 their length doubled every 188 days, and now it doubles every 121. This part would count once the length doubles every 30 days, and the doubling is still getting faster (the four parts).
• Forecasts. From the writer Vernor Vinge, the idea that nobody can predict AI progress any more. This part cannot be scored yet, because it needs a record of old predictions set against what then happened, and that record has not been built.
• AI doing the research. From the mathematician I. J. Good, the idea that AI ends up doing most of the work of improving AI. Anthropic, an AI company, says its own AI leads 26 percent of its AI research. This part would count once AI does half of that work, and only the company can check its figure.
• The economy. From the economist Robin Hanson, the idea that the world's economy starts growing ten times faster than it ever has. It grew 2.9 percent in 2025, and this part would count once it grows 30 percent in a single year.

Where the Numbers Stand

The test's answer, as of 18 October 2026, is "Not yet." All four parts would have to cross their lines and stay across for a year, and none of them has.

A plain "not yet" says nothing about whether anything is moving, so the top of the home page also gives the average of how far the three measurable parts have come towards their own lines: 32 percent. The pace is 43 percent of the way. AI's share of the research is 52 percent of the way, because 26 is a little over half of its line of 50. The economy is at 0 percent: growth of 2.9 percent is below the economy's own long-run average, so it has come none of the way (the four parts). The economy is always named beside the average, because the test needs every part to cross, so the slowest part holds the answer back. The 32 percent is not a forecast. With the economy at zero, the test as a whole has not begun to pass.

If current trends simply continued, AI's share of the research would cross its line in 2027 and the pace somewhere between 2028 and 2048, but the economy would not cross at all (the four parts). On current trends, then, the test is not met.

The Most Interesting News So Far

The pace part has lost its source. A check of the sources on 16 October 2026 found that METR's page of measurements now says it is no longer actively updated, as of 8 September 2026. METR also says its tests cannot measure jobs longer than 16 hours, and the best AI has already passed that. Without a replacement, the pace part will soon have nothing current to measure. The test's rules say a replacement must be named first, with a stretch of time where old and new overlap so they can be compared. On 17 October 2026 I wrote to METR to ask whether a new version is planned, and if not, which public measure they would point to. No answer is on the record yet (every dated change to the method).

The economy is the part that keeps the test honest. I also ran the test on the years 1995 to 2019, years when the test ought to say no, to see whether it would give a false alarm. It did not, but only because of the economy. The pace measure only starts in 2019, so I used a rougher stand-in for it, and the stand-in would have crossed the pace line on its own in 2014. I added the economy part on 8 October 2026 because it is the only part measured outside the AI field, knowing that it reacts years after the technology and so will usually be the part furthest behind. It is what stops the test from finding a singularity that exists only on paper.

How to Follow It

Every Monday morning, from 19 October 2026, a scheduled check reads the past week from the site's sources. It can add new figures and flag things for me to decide, but it never changes how anything is scored (how the weekly check works). When it finds something that moves a number you follow, I will write it up here and say why it matters, and a quiet week gets no post. The method says how the record was built and who did what. Think a number here is wrong? Here's how to tell me, and what happens next.

What to Watch For

The biggest open question is METR's answer, or a new version of its measurements, which would settle what the pace part is scored on from here.

AI's share of the research sits closest to its line, but it rests on only two figures from Anthropic so far, under 1 percent in February 2026 and 26 percent in August (the company's page). The next figure will show whether it is still climbing towards half.

For the economy, the World Bank has not yet published a figure for 2026. When it does, the economy part will move off zero only if growth beats its own long-run average (the four parts).

Next to arrive is a fuller record of what the people who run the AI companies have said about the future, in their own words and with their dates, so that anyone can see how their talk has moved. Sam Altman, who runs OpenAI, wrote in June 2025 that "2026 will likely see the arrival of systems that can figure out novel insights" (the quote, with its source). Where a date passes, the record notes in a line what happened.

Updated 2026-10-19: I opened that record the next day, with 325 dated predictions from fifteen people (the claims page).