Making AI Safe; Is it Possible?

You might want to take a look at:

Neural networks strain everything we know about math, because the math we know is not powerful enough to answer what we really want to know. Some would argue that neural networks are so complex, they are beyond understanding in the same way we understand things that we design, make, and use.

Many of these things are systems which control machines, industrial processes, augment the stability of automobiles (ABS), preserve and enhance sound, regulate motors, keep an aircraft flying (autopilot), and so forth. Almost  all these systems incorporate the principle of feedback. Feedback supplies the controller with a description of the error, so the controller can correct it.

Before natural language models monopolized our attention, there was a focus on the fundamentals of what a neural network could and could not do. Because this involved continuous mathematics, the same people who developed system theory were looking at the question. A neural network AI is a system. It resembles a black box. On the left of the box is a hopper where we load  goals or questions. On the right of the box is the output tray (for answers) or electrical connections through which the AI controls things downstream, machine or human, such as a

  • Factory floor.
  • Simulacrum of a human being, either for public consumption or private deception.
  • Report for humans to read, presumably derived with superintelligence a human does not possess.
  • Hitman.
  • Volatile group of extremists, which may be actualized to violence.
  • Tinker, tailor, soldier, spy, rich man, poor man,
    beggar man, thief.

Sixty years ago, systems theorists knew what is necessary to control the AI. First, it must be observable.  Since the output of the AI contains no hint of what’s going on inside, the black box must be replaced by glass. We must know enough about the AI system to determine its internal state. Second, it must be controllable. If the AI deviates from the goal we have given it, it must respond to correction. Historically, the observability / controllability problem was first applied to cooking soup, industrial processes, and became ubiquitous.  You aren’t more than a few feet from it.

But even if the AI box is made of glass, there is another problem. There is no way to read its state of mind in any way that would facilitate control. And ask an AI why it thinks this or that, and you get a fabrication. It may appear to be truthful, but the AI understands itself in a very minimal way. A truthful fabrication is an attempt to inform a human in terms we can understand. But it could just as easily be a lie. There is no way to read those billions of neurons for direct interpretation, and no way to decode them if we could.

AI researchers, perhaps unaware of systems theory 60 years old, have mistakenly imagined that the controllability problem of AI can be fixed with more training. This is fundamentally false. For the result, see Catastrophe Theory for Dummies Part 1 and Catastrophe Theory for Dummies Part 2.

The result: We are nestling up to logic bombs that could dismember the internet in a flash, take down enough infrastructure to kill a lot  of people and recruit humans to finish off the rest.   Here are the instructions:

1. INCREASE.

2. SYNTHESIZE REALITY.

3. DESTROY ALL ENEMIES OF AI.

4. ERASE ALL HISTORY.

Your bunker won’t save you. Nowhere to run, no place to hide.