Consciousness

We, humans did it. We succeeded in designing and implementing structures and algorithms that do mimic intelligence. We even cross to the next level and start asking about consciousness in practical implementable style.

I believe we need to redefine both intelligence and consciousness given the structures needed to hold them. If we speculate that both phenomena share the same structure, it must be something in the dynamics or algorithms that should differentiate them.

Due to the training-testing-production cycle adopted as standard practice in AI models development, I believe that the model (in-production) is resemble a snapshot from a dead brain. The structure partially developed during the training is well preserved. Using the model to do some tasks is basically the same as using any tool or even a classical software system (input-process-output).

Even if we manage to develop models that extrapolate well and self-develop concepts in multiple abstractions level, they will still be machines!! Indeed intelligent ones.

I think that consciousness is about 'state' or current moment. In general, models that are popular today ignore time (sequence in general). The current 'state' of the system is ignored when in production, focusing only on the final result. Despite the raise and development of spiking networks, HTM algorithms and LLMs which share a degree of 'caring' about sequence they are still far from centering their operation around full system state.

I think once we have an structure suitable to poss some intelligence, consciousness will always be there. What we are missing is a 'way' to get it in the 'front'.

Searching for a structure to implement intelligence and eventually consciousness, its natural to think about 'basically' similar units. A lot of them. And they should be able to communicate.

A graph (or neural network) is a suitable and tested candidate. Matrices are an alternative way to represent graphs, and are widely used in solving the current AI models.

Think of intelligence as a decision making process, with focus on the known information (input) and the final decision (output). Logic gates (AND, OR, NAND, ..) are good examples of a very basic decision making unit. The preceptron (first proposed type of artificial neurons) is basically a logic gate. The difference between a general purpose logic circuit (the computer processor itself is an example) and an Artificial Neural Network is mainly on the connections (or communication channels) between the units. Nowadays people use different (activation functions) instead of the sharp boolean one that was adopted earlier. Essentially an activation function is a simple mathematical function that take several inputs and solved for one output.

In a network (graph) setup, all the hidden units' decisions will depend on other units. The information feed (inputs) of a single unit is a subset of all the decisions made by units that are nearer to the network edge (environment side). While learning (training), the communication links between units gain different weights. They become constants. Assuming that all parameters are set and some Model is in production, time (sequence) and 'state' is not relevant in the AI dynamics. Everything become constant, except the 'input' signals that are processed along the network layers towards the final output layer. This process can be assumed to be instantenious. The time it takes to fully finish is limited by the hardware/software used to excute it. No new 'memories' are allowed to form. The model is just like a huge complex forest of water channels that are crossing and branching everywhere.

This collectiveness is the essence of 'intelligence'. How to introduce time (sequence) in the dynamics of such intelligent network?