AIQ, Round 3

For the new book I'm making several passes at describing intelligence, both human and artificial. This essay is a mile-marker to show where I currently am.

· 4 min read
AIQ, Round 3

There are three kinds of intelligence:

  • Intelligence-within-self: when a thing needs no signs or symbols at all to accomplish the widest diversity of behaviors possible
  • Intelligence-to-self: When a thing creates, modifies, and consumes the minimal amount of external signs and symbols needed to accomplish the widest diversity of behaviors possible to increase intelligence-within-self
  • Intelligence-among-selves: When a thing creates, modifies, and consumes the minimum amount of  of external signs and symbols needed to accomplish the widest diversity of behavior possible to increase intelligence-to-self

This means there are several terms that need consideration:

  • AIQ: Artificial Intelligence Quotient, a universal and artificial way to evaluate the intelligence of any thing or group of things
  • Signs and Symbols (SAS): any form of communication, from a river dance to a downloaded movie. Measured ultimately in bits. I  believe that work has already been done and is not included here.
  • Behaviors: Any physical interaction with the environment outside the thing. May be observable or not. Oddly enough, it is impossible to describe behaviors without using Signs and Symbols, although it is possible to mimick them
  • Widest Diversity of Behavior Possible (WDB): Anything that changes outside the thing being considered caused by that thing
  • Creates, Modifies, and Consumes Signs and Symbols: This is the evolution of language, as studied by linguists. Note that writing is a very, very small part of language (Perhaps less than 1%?), many liguists do not consider writing as a part of language since it actively prevents all three of the above qualities, and that language has repeatedly developed, evolved, and thrived in the absence of writing for tens of thousands of years. Mimickry is the foundation of any language.
  • Widest Diversity of Behaviors, Internal (BI)
  • Widest Diversity of Behaviors, Paired (BP)
  • Widest Diversity of Behaviors, Social (BS)

Using simple multiplication and division, and minimizing and maximizing the various critera as describe above, we get:

AIQ = (BI * BP * BS) / SAS

There are a couple of really wild things here.

  • This works for anything from a function call to a supreme being.
  • Without observability of behaviors, there's nothing to be calculated, a tree falling in the forest and all that
  • A thing can only be as smart as you can observe it to be. To a dog, humans are gods.
  • This means that intelligence always subjective to the evaluator
  • There are different ways to evaluate the different kinds of behaviors, BI, BP, BS although the evaluation process should be similar for all
  • Because of that, although it uses symbols and math, it can only be a rough approximation of the underlying reality. Perhaps it should rather be called "AIQ for me"
  • Most importantly, none of above caveats hold true for formal systems, like (some kinds of provable) computer systems. You can measure AIQ directly for technology systems under certain conditions. Properly constructed computer programs definitely have an AIQ, wether they're part of a guess-a-number CLI program or a massive worldwide super-being
  • More cool, for the technology where you can't measure AIQ, such as deep learning or poorly-written code, you can still measure AIQ using a combination of the two systems
  • To a good degree, then, we can measure artificial (and perhaps alien) intelligence using the same overall process that we would on the biological creatures we're already familiar with. We can also handle cases where there's a little of each

That leaves us with three remaining questions, two of which involve boundaries:

  1. How do we find the scope of what we're evaluating for non-computer-programs? How do you tell one thing, organism, program, or whatnot from another one? Specifically, how do we evaluate the behaviors at each level we've outlined? (BI, SP, BS). Is a human an atomic organism, or is it trillions of individual cells? Quadrillions of chemcial bonds?
  2. How do we scope out formal, provable systems, one from another? How do we define the boundaries of system for everything that is is part of a formal, provable system? (This is probably going to boil down to the coupling between chunks of code and side effects, since behavior is the key thing we're looking for.)
  3. When we have several people observing, how do we come up with an agreed-upon AIQ for #1, or the "AIQ for me" problem

I think the biggest remaining item is answering the boundary question in a way that works for both biological and formal systems. The answers don't have to be exact since we're not trying to come up with the answer to life, the universe, and everything, we're simply trying to come up with a self-consistent system that allows us to talk with consistent terms about learning and intelligence.

With the answers to those three questions, that gives us a good foundation to talk about critical path learning at scale, which is the topic of Info-Ops 3.


P.S. Apologies that this essay may seem like a repeat. I set out this morning to write about The Argument Game, a way for groups of people to reach useful conclusions to difficult and sometimes intractable problems. Instead I ended up back elaborating foundations. I find that fascinating, since it tells me that my foundations are not in place enough to move into organizational learning. We're close, though.

In brief, to tee up the next essay, arguments are simply the exchange of signs and symbols at a social level about and including a wide variety of terms, i.e. signs and symbols. To increase intelligence, to learn, we want to minimize the denomiator of the above equation (SAS) while maximizing the numerator (BI * BP * BS). This should be the goal of the Argument Game. Ideally you find you disagree about nothing in which case the demoniator is zero,  the result goes to infinity for that particular system and you're done. You never have to revist that term for that particular system. The only other successful end to the game is finding one and only one term that two groups disagree on where 1) all of the supporting terms are in agreement between the two groups, and 2) agreeing on defining that one term has the maximum impact on increasing the numerator, behaviors at all three levels: BI, BP, and BS. We're going to do this by stacking negative space, which we explained earlier.

The next Info-Ops essay should start with boundaries for computers, biological, and mixed systems and end with The Argument Game. (knocks on wood)

Related Articles

The Canyon
· 4 min read
MLS Theory and Language
· 7 min read
MAG-Lite
· 1 min read
Info-Ops 3: Introduction
· 13 min read