AI Brains Will be Analog Computers, Of Course
December 4, 2014
That same interview with Jaron Lanier had a whole bunch of insightful comments by famous futurology-type people, and I thought one in particular was excellent. George Dyson pointed out something that should be totally obvious if people would only think about it – that a true AI machine that really thinks for itself can’t be a digital computer, but it must be an analog computer.
Here’s a quote from the actual comment that Dyson posted:
The brain (of a human or of a fruit fly) is not a digital computer, and intelligence is not an algorithm. … The brain is an analog computer, and if we are going to worry about artificial intelligence, it is analog computers, not digital computers, that we should be worried about.
Isn’t that obvious? Our brains are not digital, so why would we expect true AI machine brains to be digital?
Dyson also points out something less obvious – that intelligence is not an algorithm. He’s right on this point too because intelligence means the running of an algorithm in pursuit of a goal. There’s a big difference between an algorithm written down somewhere (sitting inert) and an algorithm in action, with the juices flowing through it.
It’s the same difference between a picture of a bird flying and an actual bird in flight. Or between a traffic light turning green and the traffic actually going. Vroom-vroom! It’s the difference between a symbol and the thing symbolized. Or as John Searle would say, it’s the difference between syntax and semantics.
The point is that we’re never going to build true AI by writing down instructions for AI behavior. The only way to succeed is by building an AI machine that will really behave, in the real world. Not instructions, but action.
How can we build an AI machine that acts on its own? First of all, it can’t be a conventional algorithmic computer program, but it must be a neural network. And it can’t be a neural network that we purposely train to serve our human purposes, but it must be a neural network that evolves by natural selection to pursue its own purposes.
In another comment, Lee Smolin asked, “Is there any concrete evidence for a programmable digital computer evolving the ability of taking initiatives or making choices which are not on a list of options programmed in by a human programmer?”
It sounds like Smolin is just waiting for it to happen, as if a computer can just start evolving all by itself. And that is never going to happen. In order for computer evolution to start, we need to set up the evolutionary environment in the computer circuits, and we need to spawn the first generation of neural network threads that will compete with each other.
So I’d like to rephrase Smolin’s question. Instead of asking whether it has happened, I would ask whether we have made it happen.