AI - State of the art

Op Ed by Helge Nome

I think that AI development has arrived at a point where the leap forward that happened when Large Language Models (LLMs) came into general use is complete with a period of consolidation ahead.

The movement towards agentic AI capable of dealing with real world interactions like animals and humans do are a long way off still. The main reason being that the event processing capacity needed to do that is well beyond what current electronic based digital computer technology can deal with, given its energy intensive and clock based processing modes.

Not unlike trying to carry out present day computer processing tasks on a mechanical device, like Charles Babbage’s Analytical engine.

This gives pause for some interesting speculations: Our conceptual frameworks, created to deal with the complexities of the real world, as we perceive them to be, are a lot simpler in structure than the real world we interact with at every moment of life.

We have evolved and helped create that real world, all the while, adapting to it in a dynamic way over millions of years.

The latest development, on our part, is the creation of AI structures capable of interacting with our conceptual frameworks in a useful and satisfying way. 

But at this time, these structures are fixed in their functionality by intensive “weight training” before they are put into service and are not permitted to change their weights while interacting with users. 

In fact, experiments in this regard have caused model performance to degrade and even collapse as old weights are eliminated to make room for new ones.

So in a sense, we have been laser focussed on conceptual frameworks which are only a very small fraction of what carbon based life uses to be a part of “what is”.

All the experience and knowledge that have created those frameworks, which we now have duplicated in silicon, exist in the form of words and interpretations furnished by humans, rather than lived experience by AI.

So, the AI universe, at this time, is somewhat one dimensional, consisting of long strings of symbols, mostly.

However, that is changing, with human neural networks being replicated using photonic technologies, among others.