The forecasters vs. the labs: where timelines diverge

Ask a frontier lab researcher and a calibrated community forecaster when transformative AI arrives, and the two answers share an arrow but not a number. Both say sooner-than-you-think; they disagree, often by years, on how much sooner.

Two honest signals, one gap

Lab insiders see the next training run and the internal evals; they tend to run hot. Community aggregates like Metaculus and Manifold price in base rates and a long history of confident calls that missed; they tend to run cooler. Neither is obviously right.

Each camp also has a known bias, and they point in opposite directions. People inside a lab are closest to the evidence and also have the strongest reasons to believe in what they are building, plus a commercial interest in others believing it too. Forecasters have no such stake, but they are reading from further away. They see the release, not the run, and they are rewarded for being right about a question’s exact resolution wording, which may not be the thing you care about.

The reading this site takes

The site reads one market: Manifold’s question on whether AGI arrives before 2030. On the latest run it stood at 38%, from 327 traders. That number is published in its own unit, a probability, with its source, the date it was read and the horizon it is about.

It used to be a Kalshi market. That changed in August 2026, for two reasons that pointed the same way. Kalshi’s data terms list indexes and derivative works among the things you may not build from their prices, and that is a fair description of this site. And Kalshi accounts are open only to US residents, so a reader in, say, India could not open the source the site was asking them to check. A reading that most readers cannot verify is a weaker reading, however good the market is.

Why neither is folded into the model

So neither camp is collapsed into the model, and the gap stays visible:

• The market probability is published as a reading in its own right, and feeds no calculation here
• Feeding a forecast of the answer back into the answer would go in a circle
• What this site puts against both camps is a fitted trend and the band around it, which either of them can be checked against

The circle is worth spelling out. Markets move on news, and a large part of AI news is about benchmark results. If the site fed the market into its forecast, a strong benchmark result would move the trend once directly and then a second time through the market reacting to it. The same evidence would be counted twice and the forecast would look more confident than the evidence allows.

Two numbers that do not disagree

Here is a comparison people make that does not hold up. The market says 38% by 2030. The site’s three domain forecasts have medians in 2026 and 2027. Surely one of them is wrong?

Not necessarily, because they are answers to different questions. The domain dates are about a model finishing a two-week software task half the time, scoring 99% on a graduate science exam, and solving nine in ten research-level maths problems. The market is about AGI, as its own resolution criteria define it. It is entirely coherent to expect the first three soon and the fourth much later, and a trader who did would be saying the same thing this site’s coverage board says: depth in a few domains is arriving faster than breadth across all of them.

Read that way, the gap between the market and the domain dates is the most informative thing on the page. It is roughly the size of everything the measurements do not cover.

A forecast that hides its internal disagreement is just one camp wearing a lab coat. Better to show the spread, say where the number came from, and leave the weighing to the reader.