Peer-to-Peer Super Intelligence

Imagine you have running on your computer a small piece of code, even something as small as 1 kilobyte. Consider this small program a node. Imagine this node in turn generates output data that is based on two things - the first being determined by the specific algorithm its running and the second the data it receives at any one instant from hundreds or thousands of similar nodes elsewhere on the network (i.e. nodes running on other peoples computers). Imagine still further that all of these nodes, potentially hundreds of millions of them around the world, are all running and communicating with each other at light speed, and at increasingly faster speeds, based on available idle cpu time. At first you might compare such a scheme to SETI@home or other similar grid computing projects. The difference here though is that unlike grid computing projects, these nodes have no central controlling authority. This is important, just as the brain has no central controlling neuron or clusters of neurons.

The next question is what happens next? I'm not entirely sure, but the possibilities are both compelling and disturbing. First of all, it would depend on exactly how the seed node was created and instructed. What if it was given the simplest of rules, just like basic cellular automata that Stephen Wolfram describes? In this case, we would have no idea what evolve until it was released and ran through potentially quadrillions of massively interacting iterations. However, unlike basic cellular automata which occur on 2D or 3D grids, this p2p node system would, like the human brain, interact with tens of thousands of potential "neighbor" nodes. Further, these nodes would not just interact based on locality (latency times), but on semantic connections as well. The possibilities for experimentation here are endless, and even risky. Why risky? Because the level of complexity this p2p node network could achieve would be astounding, possibly exceeding the complexity of the human brain by orders of magnitude. It would represent a major advance over traditional neural networks (like the human brain) because connections would be made based on any definable means whatsoever, rather than on neurochemical convenience or traditional computer architectural constraint. The constraints would be more software than hardware based. And since computers are general computing platforms, such software constraints may not exist as we currently know them, since this type of informational exchange would exceed and surprise the person(s) who created the original seed node in the first place.

What this would essentially represent is a totally decentralized, out of control seed based cellular automata intelligence.

To me it seems the creation of such a thing would be totally reckless. Who is to say someone hasn't already done something like this? Who is to say that right now, running on your computer is some trojan node doing just such a thing?

Amazingly enough, what I have just describe would be one of the simplest possible implementations, and yet it is quite possible the end result of it running long enough could be some kind of super-intelligence. In other words, there might be a superintelligence running on the internet right now. One of many problems with this type of superintelligence is it emerged based on totally self-referential internal feedback. Nowhere in the scheme were humans, or the human world even taken into consideration. Quite possibly this type superintelligence, assuming it could evolve from the above scheme, would have no idea that humans even exist. It would by all definitions be an alien intelligence in every way that counts.

Ok, now lets start this experiment over, except this time we take a great deal more care about how its created and implemented, starting with the first node. What if we created a node that was general purpose enough, but with the specific goal reflected as much as possible the needs and wants of every participant on the network. A kind of avatar or virtual assistant that learned from everything you say and do. Obviously creating such an avatar would depend on the level of software advancements currently available. Lets assume we created this node today with the best minds in the world creating the best possible virtual avatar that we can. This though is only half the problem. The other half is how to these virtual avatars (smart nodes) interact with all the others. It seems that there would be a nearly endless number of semantic levels in which they could communicate. What if instead we created these avatars as basic "blanks" in which each human user defined them as they went along? More importantly they are open-sourced so that the best minds could continue to offer plug-in modules that enhanced these avatars in different ways. This way the marketplace of supply and demand of human needs would determine the courses, pathways and options these avatars took.

Meanwhile, why all these millions of avatars are running, even duplicating themselves, and intermingling at hyper-speeds all around the network, communicating with millions of other avatars, carrying on complex and subtle negotiations on your behalf, and in turn modifying themselves as they go along, what then do we get? Would we get a superintelligence from this as well? And if so, what kind of superintelligence would emerge? Interestingly such a superintelligence would likely be far more reflective of all of us. In a sense this new emergent superintelligence would be us. No longer would it simply be the amorphous global brain of all of us talking, but an actual instantiating global superintelligence that continually learns and improves upon itself based on constant human input.

One possibility of course is that someone or group of someone’s could modify their avatars to subvert the system towards less democratic means, in which their avatars, like agent Smith, takes over every one else’s. Who is to say what would happen then.

As you can see, I have no easy answers to how all of this could evolve, but the possibilities of p2p superintelligences seem to be much closer than I previously realized, for good or ill. One thing is for sure, the more open sourced it is the more resiliant it becomes to malicious tampering. The question then is what can we do to create the most friendly p2p SI possible? Eliezer Yodkowsky has his own ideas, and I'm pretty sure they are not p2p based. I would be very interested in getting comments from working members of the Singularity Institute on this.

Related:

Peer-to-Peer Games
Peer-to-Peer Virtual Culture

Posted by paul at December 28, 2005 06:12 PM