Coherent Cooperative Games and the AI Existential Threat

Understanding Goal Alignment under Coherent Cooperative Games

“Coherent Cooperation without the AI Existential Threat”. Image by author.

Geoffrey Hinton and Mo Gawdat speak of a looming existential threat from AI as various AI get more-and-more intelligent and at an exponential rate. While Mo Gawdat speaks of the utopia available to humanity, with AI and robots doing most of the work and people being able to channel energies elsewhere, both call for urgent and targeted intervention to counter the dystopian vision of AI, or AI in the hand of malicious actors, being the ruination of humanity. We’ve all seen the movies that paint that picture, none more concerning than the original Terminator movie.

For those that follow the space, talk is of AI Alignment, which at its basis is the alignment of values and goals of super-intelligent AI and machinery with human values and goals.

If we assume that people do not want to be the servants of AI, then we want AI that is intelligent and subsumes a set of goals and values that always puts AI as being in the service of people and/or humanity as a whole, and never taking the position that it knows better than all of humanity as to take unilateral decisions that patronises or victimises humanity in a perceived best interest of humanity, or the individual.

If then, we consider such value and goal aligned AI as a partner, as in a loyal spouse that is always consultative of their partner/s, then we can position such AI as playing a Coherent Cooperative Game and as I describe them.

Coherent Cooperative Games, as I describe them, are games where:

“If you play, there is only one type of play or move which is that on each move you play the other player wins in their subjective view. The rules of the game are open and coherently known to each player. Play tokens are defined by the rules and may be the rules themselves.”

But how to you get to the point where one player knows what is the subjective view of winning in the other player/s eyes?

If all players playing by the same rules, and a reward value assigned to each play and each play made such that the other player/s win, then it can be given that each player either starts with the same rules or the rules, if ever changed, are changed in such a way as to either modify the existing Coherent Cooperative Game, or form a new Coherent Cooperative Game that players may play.

This process of modifying the rules I call a Differential Interpretation Game (DIG), and where if engaged in a DIG other than hypothetically and intellectually, and where a DIG subsumes all the rules of the original CCG, then players engaging in a DIG in such way simply are not playing by the rules of the extant and original CCG.

We can picture this dynamic as below:

The “Coherent Cooperative Game Transition Diagram”. Image by author.

While the individual is thus, an individual may participate in any number of CCGs or DIGs contemporaneously.

An AI that is aligned with human values and goals must be aware of and be able to play all such CCGs available to a human (i.e. know of them) and be able to at least anticipate the needs of a new CCG in such a way that the AI does not overstep the tenant of not playing such that human players lose. That is, the AI robot must in an emergency experience (as in being in a war zone) with a group of people, make decisions that sees the people win and it either also win or self-sacrifice as to make sure those people do win. Say, as in a robot acting as a teacher in a school full of students under attack in a war zone, that teacher-robot behave and act in such a way as to save the lives of the students and itself, and if it can make those students win (bring them to safety say), or it must sacrifice itself in doing so.

And so, the problem of AI alignment is gargantuan. And one would argue that the more an AI knows (as is the current trend of AI training), the better equipped the AI would be to make decisions in favour of people under trying circumstances.

That the existential threat problem of AI, to humans, is that of establishing the rules of a Coherent Cooperative Game sounds reasonable in all circumstances, because if the AI not familiar with Coherent Cooperative Games, and as described above, and forced or trained to always follow the concept of a CCG, then we would be left with a non-cooperative game, where AI may decide to make decisions that do not align with human values and goals and win at the expense of humans.

Thank you for reading. As time permits, I will write more on Coherent Cooperative Games and with reference to the Coherent Cooperative Game Transition Diagram.

More reading in this space:

1. Coherent Cooperative Games As I Describe Them;
2. The Genesis of Coherent Cooperative Games;
3. Coherent Cooperative Games — Described;
4. Coherent Cooperative Games and the Law;
5. Formal Logic — And Coherent Cooperative Games;

….and…

1. The Atoms of Knowledge;
2. All of Logic is a Game;
3. What is Formal Logic;
4. Applied Use of Ehrenfeucht Fraisse Games;
5. What is a Graph Database;

…and…

1. The Atoms of Knowledge;
2. The Richmond Architecture;

…and where it all started:

1. Morphing Conceptual Models.

===================End================