Ultimate Intelligence

[Eray Özkural]

Dear colleagues,

This is the draft of a paper I have been writing that tries to elaborate AI theory from a purely physical point of view. I am basically exploring the ultimate limits of intelligence, extending our toolkit just a little. It likely contains many errors. I would be grateful if you could criticize it so that I can improve my paper. I will try to acknowledge all comments that I can make use of. This paper was  borne ouf of the cognitively significant discussions on the MAGIC list, therefore I am looking for comments from participants on the list especially. I first authored it to counter some of the Platonist interpretations of AI that were discussed on the list. In particular, I believe that we have to consider the constraints that physics places upon our world, and we should be careful to distinguish physically plausible statements from metaphysical (meaningless) statements. Therefore, the crux of the paper is phyicalism/constructivism vs. Platonism, philosophically speaking. The other major "philosophical" concern of the paper is to abandon any irreducible constructs, such as utility functions in measures of intelligence. I did not quite accomplish that, but I gave a limited account of that approach. I was also able to approach physically meaningful metrics of sorts that you will see. In other words, I am trying to put together a conceptual framework that is properly physicalist, reductionist and finitist. Although these philosophical positions might not seem relevant to many AI theorists, and there are not any (seemingly or actually) complex proofs or theorems, I believe they are, as explained in two papers of Solomonoff about the history of mathematical AI theory (Solomonoff's work is a mathematical elaboration of Carnap's theory). Therefore, I think the paper is more about "philosophy of AI", than "mathematical AI". It is definitional, rather than investigating bounds -- which is what a typical JMLR paper is. The excursion into digital physics is quite tangential, it is present in the paper to justify some relations I considered; they might turn out to be wrong, I am aware. I realize that I might have to remove it from the final version since there isn't strong evidence for digital physics -- it might be considered metaphysical by some physicists. With these caveats, please enjoy the paper. Looking forward to your cognitively dense comments.

http://arxiv.org/abs/1501.00601

Ultimate Intelligence
Eray Özkural
(Submitted on 3 Jan 2015)

We argue that Solomonoff induction is universal and complete in the physical sense via several strong physical arguments. We argue that Solomonoff induction is fully applicable to quantum mechanics. We refute an extraordinary theoretic claim about AI which contradicts with the physical constraints discussed. We show how to choose an objective reference machine for universal induction by defining a physical message complexity for the presently considered ultimate intelligence research program. We introduce physical measures and limits of intelligence that are rigorously objective. We propose volume and energy measures that are more appropriate for physical computations. We extend logical depth and conceptual jump size measures in AIT to stochastic domains and physical measures that involve volume and energy. We show the relations between energy, logical depth and volume of computation for AI. We also introduce a highly relaxed, general physical model of computation that we believe to be more appropriate for our universe than other low level automata models previously considered. We briefly apply our ideas to the physical limits of computation in our universe to show the relation to ultimate intelligence.

Best Regards,

--
Eray Ozkural, PhD
Founder, Gok Us Sibernetik Ar&Ge Ltd.