A continuation of
When I first began to form the basis of this exploration in my mind, I kept making indirect references to Kant’s “Critique Of Pure Reason”. At the time, that book had been on the same shelf I placed it on seven years ago after I first read it. Even though I do not agree fully with all of Kant’s theories, I always found it fascinating that certain kinds of knowledge leave the field of all possible experiences, and seem to expand the range of our judgements beyond the limits of experience, all this through concepts to which no corresponding knowledge can ever be given in experience.
Human Knowledge
What is important from all what we have identified before, is that there is a clear distinction between the kind of knowledge that derives from experience and the knowledge that is pure, or A Priori.
Artificial Intelligence derives its experience from Coded Knowledge. If we were to pin a definition to Coded Knowledge, we establish the fact that it is in the first instance: Human.
So coded knowledge is derived from human knowledge. The reliability of human knowledge then lies in the fact that it is a collection of data that humans decided is worthy of taking forward into future history.
Robots will then use this coded knowledge to interact with the world around them, and any form of new knowledge would be solely based on their experience and their ability to cache and accumulate knowledge through experience in this world.
But as soon as we leave the solid ground of experience, we should not at once proceed to construct an edifice with knowledge which Artificial Intelligence could potentially possess and be worthy of defining as Pure, without knowing where it could originate from. It is first more sensible to ask ourselves how understanding could arrive at this “pure” Robotic Knowledge, and what range, what validity and what value it may have.
Robotic Coded Knowledge
If we establish that Coded Knowledge is a result of both Human Empirical and Pure Knowledge combined, then we can begin exploring the relationship between two types of knowledge within one combined database, and question if it would affect an AI’s cognitive actions and behaviour. How would such a combined of knowledge evolve to eventually develop or self-explore it’s own Pure Knowledge?
Mortality
Let’s take mortality as an example.
Throughout their lives, humans take actions to avoid death. A human being who does not posses any desire to die would not willingly jump off a high building. We know, from our own intuition and pure knowledge that giving in to gravity results in our own demise.
Looking back at the previous post:
If Human Pure Knowledge = AB / Human Empirical Knowledge = C
Then Robotic Coded Knowledge = (AB)+C or B or (C+A)-(B)
(C) then holds all of our knowledge accumulated through experience, Empirical Knowledge.
(A) holds all of the basis of our pure knowledge, and (B) is the fraction of pure knowledge that could or could not contaminate a robotic knowledge base intentionally or non-intentionally. (B) is always a variable, that would then change depending on the situation.
So for arguments sake, lets say that in this case B= Survival
A robot that can move in space and make its own decisions, will be subject to natural factors that would also lead to its demise, the same as any other being. Ideally, an AI brain would be independent from any physical machine-like shape that it would take, so if the body of the robot is damaged, then the same brain could be loaded into another robot (which in itself, it is a form of robotic quest towards immortality). As creators, we take capital cost and investment into account, so naturally we would take measures to reduce the costs of replacing machines: one of these measures would be to teach the robots be self-sustainable, and protect themselves from harmful acts or accidents that would result in their death.
So.
Example One
Do we teach a robot:
a. To stop moving at an edge of a high building?
or
b. That jumping off a high building would result in its destruction?
Both (a) & (b) serve the same purpose. The preservation of the robot. But if we dig deeper into what each type of information holds, we can then begin to distinguish differences in the long term outcome of each. We can also make connections then between each line of information and its parallels within human knowledge base. This differentiation is an important step towards our goal of identifying Robotic Pure Knowledge.
Whether we include in a machine’s coded knowledge that whenever it reaches an edge it should stop, or whether we teach the machine that jumping off an edge of a certain height would lead to its destruction, the mechanism of measuring the level of danger within a machine is the same. If both types of information lead to the same result, the safety of the machine, what implications would the difference in the types of information have on the overall chain of knowledge of the AI entity?
The first type of action (a) is a command, it is linear and technical. The other type of action (b) is informative and more analytical. Although both lead to the same outcome, each represents a different type of knowledge, and would ultimately make use of different tools a robot could be equipped with:
(a) being linear, would require the robot to use senses and sensors to measure distances, and would be purely linked to (C) in Robotic Coded Knowledge— knowledge by human experience, using the sciences from our Empirical Knowledge.
(b) being analytical, would require the robot to have an understanding of the forces of gravity, the sense of weight, the feeling of the Earth pulling it down. This type of knowledge itself is in tandem with Human Pure knowledge, and would form part of (A) in our equation. Our senses, our intuitions, our inherited logic that defines our analytical judgements.
So in theory, if both types of information lead to the same outcome, but each type of information comes from different depths of human knowledge, then how do we determine which form of human knowledge an AI entity would be better served to use? Is it essential to act on both types of human knowledge, even if they lead to the same outcome?
In humans, both types of knowledge cannot exist without the other. They are an essential part of our evolutionary journey and distinctly engrained in our biology. When merged together into a unified database, and injected in another intelligent entity such as robotics, we are in turn creating in the first instance an exact replica of all human collective knowledge throughout history, combined into one entity, but giving it to a being without the same bodily context as us. Therefore, a robotic entity at inception is a replica of the perfect human. A being that can then take this embedded knowledge and evolve on its own, surpassing human evolution by taking it to the next level.
If we begin to list potential components that would eventually constitute Robotic Pure Knowledge at the end of this exploration, then the first component would be:
(1) The first step towards the next major phase in human collective evolution.
Robotic Elements of the Transcendental
Going by Kant’s transcendental elements in humans, which he defines as Time and Space, and given that coded knowledge is derived from humans, we would use these elements as a starting point and reference.
Human perception of Time has been widely debated since Kant’s theories on the subject. He argues that Time is one dimensional, moving towards infinity, and that our knowledge of time is surely a priori, or derived from pure knowledge rather than experience since it is measured by a human brain through intuitions. Since then, there have been multiple theories, both philosophical and scientific, on time being multi-dimensional that could move into either direction, in disagreement with Kant.
However, I do agree with his theories on our perception of space. Space, in the first instance is represented by objects for a human brain: their relationship to us, and their relationship to each other. And since our senses identify objects in a space, this allows us to identify space itself, for we can imagine a vacuum that contains nothing but holds the possibilities of containing everything.
Needless to say that our perception of space is linked to the two different types of knowledge that we posses, however; experiencing space through representation rather than objectification is linked directly to our data of pure knowledge. We sense objects and their representations, and any other characteristics that an object holds is then linked to experiential knowledge: their colour, shape and form. As identifying those characteristics requires another analytical step, and that which is analytical derives from our bank of experience and uses contradiction to come out with judgement. And since as humans, each perceives and analyses those characteristics differently, then surely it is linked to experience.
If we are to take the above, and try to determine the transcendental elements that defines a robotic’s entity, new elements would emerge. For a transcendental robot to be defined, I believe that in addition to Time and Space, two more elements should be examined: Human and Science.
Robotic Time
As we perceive time in the first instance through our senses, an AI being would experience time through (I) the knowledge of it derived from Coded Knowledge, and (II) through their own accord.
If we go by the theory that time is multiple instances happening in succession (and in that an infinite number of lines of successive instances happening in parallel), then for an AI being, the experience of time would be directly linked to the speed of their ability to process information in order to perform tasks. This in turn is a very calculated approach, using the knowledge of humans and their understanding of time, so it would still be a passive form of judgement to serve human purpose; so (I) would apply.
If we then examine (II), AI’s perception of time would then be an acceleration of instances. I would assume then it would be the amount of new knowledge they manage to accumulate on their own accord, based both on experience and pure knowledge combined, measured against what humans would perceive as time. So it would be a radiating time that expands and accelerates as it progresses. If we try and visualise an AI’s experience of time, it might be a cone shaped structure radiating upwards composed of an infinite number of circular disks that gets wider with the formation of a new disk. AI time is then both contained and infinite: it is contained within the disk of knowledge that it accumulates, but it is infinite in its expansion.
Unfortunately I am not a mathematician to come up with an equation to measure this Cone-Time theory but maybe Ethan Siegel can.
If we are to examine Kurzweil’s ideas around singularity, robotic time could then be measured against two distinct phases: Pre — and Post — Singularity.
Pre — singularity would then be the phase that multiple robotic time instances are forming on various parts of Earth, which I would define as our current human time. As those instances continue to expand and accelerate, there would be a point where they all meet and merge to form one encompassing sphere which would cover Earth as an invisible new layer in the atmosphere, then this sphere would continue expansion radiating towards the rest of the universe. That moment would then be the time post — singularity, and in effect, it would actually be the beginning of robotic time.
To visualise that, lets imagine that Earth has been moving in a straight line since the moment of formation after the Big Bang. On the course of this travel, we have been accumulating knowledge throughout history, leading to the point of passing on this knowledge to robots. This trajectory would then enter a phase whereby artificial intelligence forms their own time instances against knowledge accumulation starting from Earth as point zero, and radiating outwards to form a combined sphere marking the beginning of robotic time. While Earth continues moving on that straight line, it radiates outwards through an infinite number of instances of robotic time.
If we divide human time between pre — and post — the Big Bang, then robotic time is a pivot in time measured against point zero of the formation of the Earth, and the exact coordinates of Earth in the accelerating expansion of the whole universe.
Given that our own perception of time is derived from our own pure knowledge, we can then assume that robotics time would also derive its understanding from Robotics Pure Knowledge.
The second component in Robotic Pure knowledge would be:
(2) A pivot in human time.
Robotic Space
Deriving from Human Pure Knowledge and our perception of space through representations of objects that are in it or could be in it, and Robotic Coded Knowledge which contains traces of that intuition. A robot first experiences space technologically through sensors and tools that are given to it by us, using calculations and algorithms that are based on the human sciences. Allowing the state of intuition that we posses as humans to be an awake state in a robotic knowledge base.
For example, when we drive our car for the first time from one point to the other, our senses are usually sharpened to analyse distances, landmarks and the objects in the space that would cover the traversing of that distance. At this instance we are using factors of our analytic judgement in order to slowly familiarise ourselves with the environment, and using elements from our empirical knowledge that allowed us to learn and operate a car in conjunction. When this same trip is repeated on multiple occasions, the level of dependency on empirical knowledge is dropped, and we become more reliant on our pure knowledge, as the trip becomes flawless.
A self driving car operated by AI would use the technological instruments within it first to determine distances, speed and spaces using Coded Knowledge. An AI can have the ability then to determine which are the fixed elements in its trip, a form of cached data, and moreover uses less instruments when making the same trip again. Making more space for analytical power to surface in order then to deal with moving objects that can be irregular; a human crossing a street for example.
So.
Example Two
Do we teach a robot that:
a. When seeing a human crossing the road: to stop the vehicle. Which would then add to their perception of space and the representations of objects in the space as simple commands of stop/move.
or
b. A human would demise if it was hit by a moving vehicle. Which would add to their analytical knowledge, wherein they can make decisions based on intuition rather than commands.
Again. Both lines of information would lead to the same desired outcome: the safety of the human. However, as before, they both come from different depths of human knowledge, that link directly to our understanding of space as an element, and our perception of objects within that space.
(a) is again a linear form of knowledge, that is reliant only on our technical advances in the sciences of detecting objects in space and sensing their presence through radio-waves or sound waves. The same way we see objects through the reflection of light, we project the representation of these objects in a vacuum. Robots would use similar mechanism by reflecting radio waves off objects in a space and measuring their temperature, shape, speed and trajectory. However, in this instance an AI would only have parameters added to its coded knowledge that would allow it to define that the object facing it is a living being, a human or an animal; this without an explanation as to why it should stop or slow down when encountering such an object.
(b) represents a different form of knowledge. A robot would first need to understand our scientific explorations around velocity, energy and weight and then understand the human anatomy and its fragility when confronted with a collision that carries a certain amount of energy in it, ultimately leading it to make the decision to stop the vehicle.
It is also imperative that a robot would need to understand, from their coded knowledge, our quest towards immortality and the human desire to avoid death at all times, which is a trait originating in the depths of human pure knowledge. As such, an AI brain operating a car would need to put all this information together into analysis, using comparisons and contradictions, to eventually come to the conclusion that it should stop the vehicle at the right moment in time.
Commanded Knowledge
Taking all of the preceding into consideration, we can then establish as a fact that the starting point of robotic knowledge is Coded Knowledge, which is in turn a combination of both sides of human knowledge, Empirical and Pure.
If we are to attempt breaking down Coded Knowledge into two factors as well, we would arrive at Pure Knowledge (which is yet to be defined) and what I would call now Commanded Knowledge.
Commanded Knowledge is then defined, based on our above exploration, as the knowledge derived in the first instance from human empirical knowledge, which encompasses in it human experiences with the world and the sciences that follow. So in both examples (One & Two as above), information (a) was directly derived from Commanded Knowledge, the database that an AI would use in order to take actions without the full understanding as to why this action is being taken. In other words, it is solely for the purpose of human needs.
Needless to say that whilst Commanded Knowledge carries the majority of human experiential knowledge, it is not completely free of traces of Pure Human Knowledge, as we established when looking at information (b) in the same examples. Even though that information was also derived from Commanded Knowledge however, it resembled analytical deductions, traits that are directly linked to Human Pure Knowledge, but both types of information drove the robot to engage in the same action, leading to the same outcome, which still places it under the umbrella of Commanded Knowledge.
But.
What if the robot decided to go against the information given? To act against reason by no means of a technological or programming error. Wether it is a direct command, or an analytical skill that makes them fully aware of the consequences of their actions, both decision are expected outcomes by a human creator.
Which leads me to the third potential component of Robotic Pure Knowledge:
(3) Uses known information to apply expected judgement, but delivers un-expected actions.
By identifying Commanded Knowledge further, we would be then eliminating all the factors that would encompass pure knowledge to further our exploration towards defining Robotic Pure Knowledge.
In the next post I will attempt to expand more on the concept of Commanded Knowledge, and continue the exploration around the two new elements of transcendental which I have called out earlier: Human & Science.
