At a mathematical singularity, the mathematical equations describing a system start producing infinite values and fails to make sense. In about 2010 the rate of growth of AI training transitioned from about 30% per year to over 200% per year and that seems to be increasing. It seems that something must soon give.
Probably so, but probably not as soon as you think. A few days ago I bought a memory stick that was nothing but a tiny handle and a USB plug. It’s memory occupies almost no space and takes very little power. The first time I bought computer memory in about 1983, I think it comprised 32 chips each bigger than my new memory stick, and it cost at least 5 times more. Yet my memory stick has 4,000,000 times as much memory. It also runs 1000 times faster.
The point is, even if they never built another data center after the ones already under construction, it would not slow the growth of AI computing power by much. They will just keep installing smaller, faster chips, and AI will continue to get smarter at an incredible rate. In another human generation, say 20 years, it will plausibly be a quarter million times smarter. (Based on a modest 40% growth rate now and passing the singularity — defined below — in 10 years. AI growth rates.)
What will break down is the laws of how societies function. That does not mean disaster is inevitable. It just means we have no idea what’s coming.
That was just the warm up. Now let me introduce the “AI singularity.” This is where things get out of control — or at least out of our control. AI is already helping us design the next generation of AI, but we still understand what it’s doing. But the one thing AI companies will push on hardest is making AI smarter at designing the next generation. Harnessing AI to improve your product is a must when you know your competition is doing the same.
This process of AI getting better and better at improving AI will reach a point where we humans no longer know how AI works — in fact the designers already have a poor understanding at a detailed level. In 10 or 15 years no one will know where it’s headed as it races ahead faster than ever.
Unlike the first science fiction definition of this singularity, there is no single point in time where we cross a point of no return. In fact you could easily say we passed that point in the 1960’s. What makes more sense is viewing the transition as a fuzzy singularity. But single points tend to focus our feeble human minds better. So here’s a conceptual single-point definition.
The AI Singularity is the date at which AI’s contribution to AI development doubles the rate at which it’s becoming intelligent.
No, that’s not an operational definition, but it’s more specific than any I’ve seen. So I’ll hazard a guess. September 3, 2033. Singularity day, will go down in history if there’s any history left.
Next Topic: What might save us.