In the previous chapter , we have introduced the basic principle of ACI:

Intelligent Agents should behave the same way as past behaviors which are doing the right thing. 

But where are those examples of doing the right thing? This chapter will answer this question and illustrate how it applies to both natural and artificial intelligence.

Right things certified by natural selection

In evolutionary history, the histories are strained through a sieve called natural selection, then organisms receive and remember the right things. All the actions of the ancestors of one organism are right by definition, because all of the ancestors have survived and reproduced successfully, otherwise their descendants won’t be born at all. That is exactly how natural selection works. 

Thus living organisms, as examples of natural intelligence, could inherit the ability of making right actions, and learn to behave in the right way from evolutionary history. However, the behavior within an organism’s lifetime has no guarantee to be right.

(Yes, we know many species are extinct, but that's why organisms today do not have their genes (containing information about how to live). On the contrary, every ancestor of a living organism did not die before they reproduce.)

Thus we can introduce an event-centric view of evolution:

Natural selection and evolution are best considered from the view of actual events in which organisms interacted with their environment. Inheritable information about those events is passed down from generation to generation, while each generation of organisms is a sieve, retaining information about events which contain successful behaviors and strategies in their own environments while others are removed.

Like in the intelligent agent model, it is necessary to consider all the interaction between the organism and the environment. This includes all the events from the absorption of sunlight photons by pigments in the skin, to the behavior of beavers in constructing their dams. The genetic information can be seen as a form of memory about these interactions. However, this information is always incomplete and compressed.


The true intention behind paperclips-making

ACI asserts that artificial intelligence should also be guided by precedents, not just by goals or rewards. The right actions for a paperclips making machine is not maximizing the number of paperclips, but following the precedents of making paperclips in the real world occupied by human civilization, which includes not only paperclips producing, but also local law abiding, self preservation, and resource acquisition, etc.

It is important to note that ACI is not limited to replicating human behavior, as is the case with imitation learning models. ACI is aiming to extract the true intention behind the action of producing paperclips, which can be further optimized with the aid of increased computational power. This process will be discussed in detail in chapter 4. In the next chapter, we will answer the question: what is the meaning of behaving in the same way



We can end this chapter with quotation from Ludwig Wittgenstein:

Die Welt ist die Gesamtheit der Tatsachen, nicht der Dinge.

The world is the totality of facts, not of things.


New Comment

New to LessWrong?