Verification as the key to AI

Professor Sutton of the University of Alberta has something interesting to say about the key to AI:

The Verification Principle:
An AI system can create and maintain knowledge only to the extent that it can verify that knowledge itself.

At first glance, this may look a bit like circular reasoning, in that only an intelligent system could presumably do fully automatic knowledge verification, yet that activity is also claimed to be the empowering core of an intelligent system. However, there may be another way to achieve this, by mimicking the process by which humans have created knowledge over time.

Humans didn’t appear in evolutionary history with books on astronomy already written by forebears, yet humans now understand a great deal about astronomy,  despite the fact that primitive societies were very lacking in abstract thought. The fact that early human societies were greatly lacking in abstract thought, is shown by anthropology studies that date the ‘modern’ human behavior of art,religion and burial as starting around 50k-75k years ago, whereas humans have existed since at least 150k years ago. And it also apparently took getting to the vicinity of 15k years ago to move much beyond that. So humans started their existence with a very primitive abstract thinking ability in practice at minimum.

How did humans develop the intelligent abstract thinking skills we have today? Here’s the way I see it:

-Humans evolved a certain level of abstract thinking ability in the form of language and social behavior.

-Humans evolved mirror neurons which cause one human to feel similar feelings to other humans by the automatic mimicking of body language and facial expression. This is probably also the basis of the advanced human sense of identity, and the prominent part that sense plays in social interactions.

-Humans have a tendency to project, both anthropomorphizing the world around them in general, and also seeing their own internal psychological struggles in others and the world. Just as a tiger most likely sees the world an tiger categories of prey,sexual mates,sexual competitors, etc, so also did early humans see the world in exclusively human categories. However, the vastly greater genetic abstraction ability of the human due to language and social pressures, means there is a much greater possibility for making mistakes.

-Humans exist in communities, have language and communicate new ideas about the world to each other.

-Because of social-sexual competition, and the fact that the more correct a theory about the world is the more applicable and powerful that theory is, anthropomorphized ideas about the world improved slowly over time. (more on the specific fine details and exceptions of actual human history another time)

-Thus a pressured process existed to create and refine all types of knowledge over time, including ultimately the principles of abstract reasoning themselves. I think another important aspect of this, is that the existence of mirror neurons and projection caused humans to deal with knowledge on an identity level. Thus working with knowledge wasn’t just an academic past time, it was a core survival trait for a species to whom social competition and cooperation was paramount. I think this is true even though primitive humans probably had psychological identity systems that just barely existed by todays standards.

So the keys are:

-A certain level of genetic language ability, acting in a pressured both competitive and cooperative community social environment.

-Having the trait of working with knowledge at a core identity level, and having identity be the core part of social competition.

-Repeated new imperfect use of that knowledge in new areas regardless of how inapplicable it was.

-Enough time,birth,life, competition and death to produce a cultural evolution of knowledge over a long period of time.

To create AI programmers should recreate this same process with computers to the extent that it is possible. Ideally it would be done with AI controlling robots interacting with humans and other robots in the real world. These robots would be born,live and die, have genetics and pass them on through robot sex. Dominant robots would have more sex. A dominant robot is one whose model of the world is more accurate, thus allowing him to gain greater employment and acquire resources. The robots will need a sex drive, language,mirror neurons, and both a social cooperation and competition focus. Will the creation of AI require robots to have an identity,reproduction and competition psychological structure, a will to power, to the degree that they feel impelled to compete with humans also? Well, to some certain minimum degree I think the answer is clearly yes. For example, if a car mechanic purchased a helper robot to say fetch tools and lift engines out of cars, he would clearly prefer a robot whom he could talk to at least somewhat like a human helper. The mechanics robot would have to stay as up to speed with evolving human speech as possible.

I don’t think we need to worry all that much about robots becoming dominant over humans all that soon, because at minimum we will clearly have some amount of warning time as we observe the developments of robots. And we can observe and control that development at the ‘genetic’ level of programming. So we should see warning signs well before hand if robot competition starts to get out of control.

AI will be born out of robot competition on the free market.

Here are some additional expository parallels:

robot sex/robot dominance=competition and use of open source robot code usage. First done by humans, later by auto robot transactions on the internet.

robot mirror neurons= auto robot transactions on the internet.  (yes, same thing as sex. robots are funny that way. This equivalence also has parallels with combinators in programming.)

robot birth,life,cooperation,competition and death=the life cycle of the use of the above code. Since I am a lisp programmer, code to me also means data.

robot identity=after a beginning period of direct human selection of code, code packages will evolve into full fledged robot personalities. Human customers will try out new personality/identity packages on a regular basis, thus providing competitive selection pressures. After human customers are done with a robot personality, they may chose to wipe the newly evolved personality, or let it fend for itself on an internet free market. Thus we may some day be getting cell phone spam from orphaned robots talking like used car salesman in an attempt to get downloaded and installed into mechanical bodies.

Udate: I realized I was lacking a well explained robot parallel with this human knowledge trait:
-Repeated new imperfect use of that knowledge in new areas regardless of how inapplicable it was.

The way to do that, is to let robots both apply transformative code creation and to try out each others code borrowed from each other over the internet. Thus you may say to your robot ‘go make the bed.’ Assuming the robot already has the language ability to ‘understand’ this command, he goes into the bedroom, tries various code based attempts at making the bed, and compares them to his language notion of what a made bed is. The code base he searches through is not only his own code, but code on the internet, and also new code that he invents with by applying transformative code to old code. He not only carries out this process with the making of the bed, but also the evolving creation of his ‘notion’ of what a made bed is. The ultimate verification is his human customer expressing approval. Over time robots will become better at making beds than humans, because they will discover new ways of making beds that not every or even any human knew about it. Transformative code could be made with analogies, trial and error experimentation starting from library knowledge and probably other methods as well. For example, the robot may have a library of fabric material properties and manipulations, and use that for steps in experimentation with both napkin and bed sheet folding, and recognize the analogy between the two as fabric materials. This suggests another thing robots will need that humans have-a basic genetic ability to recognize and manipulate classes of objects in the physical world.