Communication, natural language and artificial intelligence

Natural language is about communication, and the first question that come with communication is : why we communicate. Obviously, we will answer this by two points: sharing information and asking information. I will prefer to speak about being part of the reality and testing the reality.

Testing the reality is related to the question process. Our internal system is driven by his own need, the balance between satisfaction and dissatisfaction. Each time we have to answer this, an internal process occurs going through our brain and activating neural networks equivalent to already experimented situation (http://www.grdprocess.be/Public/?p=340). But we have to align the current environment on the internal rebuild scenario to ensure an efficient result.

A situation is related to objects and their transformation trough time. External stimuli don’t give the cause and the purpose of a situation, it gives only the fact. But because the external stimuli are processed in the brain and linked to our homeostasis, we are putting context on it. The context come so far from the internal state at the moment of the experimentation of an external fact. There is, this way, a difference between the fact coming from outside and the context coming from inside. When you see a lion running in your direction, you feel uncomfortable. The fact is the lion running from A to B. Discomfort come from inside you.

The fact doesn’t give you the cause and the purpose of the lion running to you, it is the discomfort that will contextualise the fact.

But a lion running from A to B (even if you are the B) will not have always the same cause and the same purpose. This case is a bit trivial as there will be first an immediate reaction and you will not ask your teammate to know more about the situation. Let’s say the situation is unclear. It means that the external stimuli and your internal state are not aligned. Then the brain continues to process, which is felt as having questions. Having question become so as testing of the reality to align fact with internal state. It can be see as a switch. Until there is not a minimal alignment, the switch for reaction is put off.

Asking a question is so related to the link between fact and context. If you don’t know what’s a lion, you will ask about lion, because you cannot align something you have never experimented before to your internal state. It is a first experiment. But you can also ask if the lion is a danger of not, which mean again an alignment process between fact and internal state.

This means that in every case, the question process is related to an alignment issue, not a fact or an internal state. Question process is also related to what’s in your mind at the moment, which is a current situation, or , if you are closing your eye in the middle of nowhere, which is about reinforced neural path (something that is often processed in your brain and, as the brain is a continuous processing machine, when the process capacity is not enlisted to present issue, it goes where the networks is the strongest, which means the most used and reinforced or the most recently used).

Sharing information, the second communication process, is about being part of the reality. It is about pushing our internal state to the outside to model the reality to be aligned with our internal state.

Having a dinner or going to a party where we speak with other people is about having pleasure, which is an internal need. Going to a debate and present point of view is about forcing the environment to become beneficial to us (whatever is beneficial for us, it can be positive or negative for the others – it can be helping others or exploiting other, or even having no effect on others).

In both communication process, the cause is the internal state, and the purpose is answering the “will” of our internal state (http://www.grdprocess.be/Public/?cat=56).

There is a third communication process, but this one has no purpose at all, it is caused by the fact that as object, we are transmitting information like any object in the universe.

These view of the communication process is about the fundamental of communication that’s used by any kind of living form.

As human being we have developed specific communication means that are unique as the language whatever is the form of the language (the writing, the music, ….)

The difference between us and other living forms is not a question of process but a question of level and complexity of process. I will not address this point now.

If we consider the communication processes we saw, then we see that for an artificial system, there will be a missing element with the internal state. An artificial system is not a construction of billions of internal interactions, it is a unique system, as a unique cell. The only state such a system can have is about having enough energy. There is there again a difference on a time scale. The energy level of an artificial system is something that is managed on the long run, it is not a real time balancing system as a biological entity.

This way the internal state of an artificial system will be extremely poor, and a complex communication system cannot emerge spontaneously from it. We are forced to code rules to force an artificial system to communicate.

Two situations can be the same with a different context, or two different situations can have the same context, and, in both cases, a biological entity will start an alignment process. This is possible because of the biological architecture of the entity. An artificial system doesn’t have this kind of architecture. This way we can see that using natural language to speak between a human and an artificial system is an issue without putting rules in the code. But at the end, putting rules in the code will not give the understanding of a context to an artificial system.

A solution will be to explain each time for each situation the context to the artificial machine. But as a context can be lived in a few second, it can take several minutes, event hours to explain, because a context is the result of a lot of past experimentation.

At the end, for the artificial system, each combination of situation and context related to several objects (it means that a situation for mean doesn’t have the same context that the same situation for you) will become a memory and processing issues even with the technological development, when for a human, whatever the fact and the context is, we are taken the nearest experiment we have done and move to the new one.

We can also see the limitation of natural language between humans and all the misinterpretation that come from it. A philosophe a said one day that we need to speak about facts and if we are speaking about something else it is better to shut up.

Maybe we have to work in another way to try to solve (or at least decrease) the issue of communication between artificial systems and us. As implementing an equivalent to an internal state is something difficult (not far from the impossible), we have to modify natural language rules and be able to develop a factual language. This way, we will not align artificial systems on humans, but we will align human on artificial systems.

What about the context then? Using natural language to explain a fact is in my opinion enough now. But we have to develop a standardised language to explain context. The artificial system will have to ask questions about the context and we have to answer it in a logical language, so the artificial system will be able to align a context on a fact.

It is a important issue, as understanding a context will drive an artificial system to understand thinks that are not obviously factual as “human need to eat or they die”, but also to understand thinks like “human fighting each other’s” or “someone plundering resources”, so the artificial system will be able to manage situations not only by erasing all parties to stop the fight or the plundering. It goes much furthers than implementing rules to avoid negative impact on human being or to have system that will replicate human bias as we have already experiment with some agents from Facebook of Google.

To summarize what must  be done to allow an efficient communication between humans and artificial systems we see that we need to develop a formalised contextual language on our side, to train people to communicate this way with artificial system and to implement into the code of artificial systems the obligation of understanding the context and to align it to the fact by asking questions that human have to answer in this formalised language.

GRDprocess is doing research on the thinking process and its implementation for artificial general intelligence to propose track for solution.

Insight for the development of an AGI

Intelligence is a tool resulting of evolutionary process. Intelligence have two dimensions: the capability of representing the world and the capability to use this representation for effective decision making, as decision making is the key point to ensure the existence through time. The level of intelligence is a product of these two capabilities.

Working on these capabilities to build an artificial intelligence lead to two results. Depending on the development, a partial one will lead to ASI, artificial specific intelligence, a complete one will lead to AGI, artificial general intelligence.

Human being has his own representation of the world based on set of input received and experiments made.  Human being made automate computation on this world representation to take decision, but also made computation to upgrade his world representation by a continuous improvement process, there is thus a feedback process between capabilities.

Human being is a life form, as all life form, it has no purpose in itself. Life is evolving through time and, most of the time, processes maximizing existence through time are selected. Intelligence is one of these processes. Life can be seen as a chain reaction that maximise his existence through time by favouring processes that notably increase decision making, as intelligence.

Developing an artificial intelligence is thus like developing an artificial tool. It is not taking into account a purpose. It can be used by us to enhance our capability to exist trough time. Giving a purpose to an artificial intelligence (by coding a target to reach) will be as telling a human to learn more, to focus on observation, to use more computational model, … In other words, to optimise the tool of intelligence but not to give a purpose to life itself.

There is a lack of understanding about a concept called the singularity. The concept of singularity takes into account independence and autonomy of an artificial system having a capability of representation of the world, a capability of compute on those representation and a feedback process between these capabilities. What’s missing to complete the concept of singularity is what I call a motivator. But as there is no purpose in life, to translate no motivation into a motivator to be able to add an artificial system in the chain reaction of life is an issue.

We have now several points to investigate in order to design a complete process for pushing an artificial system into the chain reaction of life.

Representation of the world: the question is on what element do we have to build this representation. Knowing that each life form has his own representation depending on the way its capturing information. Human five senses are a good approach that can be upgraded. Knowing also that the representation of the world is a partial representation of the reality in four dimensions (a representation takes into account physical items into space but also evolution of items into time). The issue can come from defining up to what extend do we need to go. It is a trade-off between exhaustivity, use of resources and speed of processing.

Use of the representation of the world: to have a computable representation of the world, there is a need to transform it into concept, schemes, hypothesis. It means meanly two points: fill in the blank without having direct information and manage the causality process. All these computation process will have as target to maximise the efficiency of decision making. Because each decision will have an impact on the next move. Again, there is a trade-off to do between maximising the choice and the reaction time, there is a risk to fall into an infinite loop or to do series of wrong move which will reduce the choice and effectiveness of the system on the long term.

What’s the ultimate criterion to follow to ensure the development of a real efficient general system that will continue to exist trough time by maximising its ratio: number of choice / resources used.

In order words, what’s the motivator to implement so that when we will push the start button, the system will be totally autonomous and evolve on the long term without going into a dead end.

If we are able to answer these three questions, then we will no more speak about AGI as it is only a tool but we will speak about a new life form, something totally autonomous being part of the chain reaction of life. That’s in my opinion the concept of singularity that need to be developed.