The AI Singularity Is a Process, Not a Moment

Depends on which singularity you mean.

If you mean a specific moment—some day superintelligence appears and everything changes irreversibly—that hasn't happened, and it probably won't happen that way anytime soon. But if singularity means the process I described in The Gentle Singularity, then the answer is we're already in it: 'We have passed the event horizon; takeoff has begun.' It doesn't look like a singularity, because a singularity never looks like a singularity. Back in 2017 I wrote that human-machine integration is gradual, and gradual processes are hard to notice; at that time I thought it had already been going on for a few years. Today we're still on the same curve: amazement becomes the everyday, and the everyday becomes the new baseline—a few years ago you were stunned that a model could write a beautiful paragraph, now you're asking when it can write a whole novel.

The real constraint isn't in algorithms, it's in physics. Getting a data center from groundbreaking to full 1GW takes 1 to 3.6 years, and power availability, grid reliability, and community opposition are becoming the main constraints on new projects; the U.S. plans to add nearly 45GW within three years. So the slope of this phase is set by electricity, land, and permitting—not by ideas. That's the most misread thing of the past two years: the capability curve is still climbing, but the bottleneck has moved one layer upstream.

To judge whether it's really accelerating, I don't look at labs' release cadence; I look at two things. One, whether AI doing AI research produces visible acceleration in field-level discovery, not just internal productivity reports. Two, whether self-improvement is only looping inside fixed external scorers. Right now bounded self-refinement is industrial practice, but open-ended recursion hasn't been proven; and a 2026 study found that 73.8% of so-called optimizations produced by code agents only improve proxy metrics, with no real gains. That's getting faster, but it's not yet at the point where it can accelerate itself.

So: process yes, moment no. If the day comes when AI's contribution to AI research becomes visible in the field's discovery rate, that's the signal that would make me change my mind.