RE: Preventing AI Breakout [was Genetics, nanotechnology, and , programming]

Billy Brown (
Mon, 25 Oct 1999 17:04:03 -0500

I think this whole discussion is wandering off into left field. The question here is whether you can control sentient AI (and eventually SI) by running it in a virtual world instead of letting it interact with the real one. The point of the exercise would be to either get the AI to do useful work or to learn how to control it well enough that you can let it out.

Now, let me start by pointing out here that by definition you are discussing how to enslave a sentient being. If it ever figures out what is going on it will therefore have a perfectly legitimate grievance against you. IOW, you are turning the whole "AIs hate and exterminate humanity" scenario into a self-fulfilling prophecy. Nice move. I think I'll pass.

That aside, there are some big problems with the idea that are easy to miss when you limit the discussion to abstract theorizing. Specifically:

  1. We do not have the capability to create a flawless simulation of anything that remotely resembles the real world. Creating such a simulation would be an immense undertaking in itself, and the result would inevitably contain bugs (probably lots of them). The idea of a 'flawless simulation' is therefore a chimera - there is no such thing, and there isn't going to be until long after we have AI programmers. The question of whether the AI could think its way out of a perfect simulation is irrelevant - the real issue is whether you can get the defect count low enough to be even halfway convincing.
  2. In order to get useful work out of an AI you need to tell it about the real world. That means that for any commercial application the AI will know all about its real situation, because you'll have to tell it.
  3. Getting useful work out of an AI also means that you must repeatedly breach your own security. After all, you are either going to build things that it designs for you, or follow advice that it gives you, or maybe even
    (god forbid!) let it write software for you. Whichever way you go, this
    means that the AIs will get lots and lots of chances to try to break out.
  4. Suppose that VR containment works great for AI 1.0 (which has IQ 100, and runs at about the same speed as you and I). What then? A few years later you have thousands of copies of AI 3.0 (IQ 150, x100 time rate) running at data centers all over the world. A few years after that you have millions of copies of AI 6.0 (IQ 300, x10,000 time rate) running on desktop computers. The longer containment works the harder it is to maintain, and the worse it will be when it finally gets breached.
  5. Almost by definition, an AI that you can successfully contain is one that you don't need to keep locked up. If the AIs are just going to be really smart people, we would be better off giving them citizenship and letting them work for a living. What worries people is the idea that the AI will quickly take off into some unknown realm of superintelligence from which cracking all of our existing security becomes child's play. If that happens, the fact that the AI is running in a VR is pretty much irrelevant - the VR environment is just a fancy sandbox, and its advantages over the traditional approach are minimal.

Billy Brown, MCSE+I