Fwd: Looking in the wrong places

This article gets at the heart of the problem of what exactly is artificial
intelligence and what it is not, cannot or never will be.

https://arstechnica.com/ai/2025/07/agi-may-be-impossible-to-define-and-thats-a-multibillion-dollar-problem/

Milton Ponson
Rainbow Warriors Core Foundation
CIAMSD Institute-ICT4D Program
+2977459312
PO Box 1154, Oranjestad
Aruba, Dutch Caribbean

---------- Forwarded message ---------
From: Milton Ponson <rwiciamsd@gmail.com>
Date: Wed, Jul 9, 2025, 15:07
Subject: Looking in the wrong places
To: W3C AIKR CG <public-aikr@w3.org>, public-cogai <public-cogai@w3.org>, <
paoladimaio10@googlemail.com>, Owen Ambur <owen.ambur@verizon.net>, Dave
Raggett <dsr@w3.org>


The last few days comments have surfaced again about the data peak problem,
and the implications for training LLMs that are becoming ever larger and
ever more power consuming and cooling water wasting in the data server
farms housing AI.

At the same time research seems to point out that the generative LLMs are
not decreasing their hallucinations and seem to be not improving.

Which got me take a look at how we set up the cognitive architectures and
topologies that define the data structures, types of neural networks and
the algorithms used.

Cognitive neuroscience uses a.o. fMRI, EEG and other techniques to catch
brain activity in action. Plenty has been published about physical
processes in the brain which are both local and multi-regional,  analog and
seemingly quantum.

As a mathematician  I cannot but wonder if we are interpreting what we see
the wrong way.

The decades' long bet between Koch and Chalmers on consciousness was not
settled and plenty of theories are out there which focus on cognition and
consciousness of which I find the "enaction" and "free energy principle"
intriguing.

As a mathematician I am continuously bumping into research articles
heralding the imminent advent of a grand unifying theory or framework that
will solve some major problems.

If we take a look of our evolution as an intelligent species, the
biological hardware we use for cognition came about in a non-orderly
fashion, in some cases ancient viruses contributing to our neuronal
communication capacities leading to the development of consciousness
through the Arc gene.
From genomics we know that most of our DNA does not code directly for
proteins, but contains deprecated genes, dormant genes and code for all
manner of yet to determine (regulatory) functions.

This should tell us that the human brain, though highly developed,  well
studied in terms of cognitive architecture and topologies,  is a mixed bag
of tricks.

And the one thing that eludes us is where actual bits of information are
stored.

Penrose and Hameroff with their microtubules theory espoused a highly local
theory, yet we know certain mental processes seem to originate from
specific areas and some cognitive processes activate a multitude of regions
in which it isn't always clear whether purely analog processes are at play
or local quantum effects play a role as well.

Which means that both the cognitive architectures and topologies aren't
clear and well established yet.

Which makes the discussion about what constitutes knowledge and worse
consciousness almost impossible. And we haven't even touched upon causality
and the spatiotemporal aspects.

Which brings me to the following article

A geometric link: Convexity may bridge human and machine intelligence
https://phys.org/news/2025-07-geometric-link-convexity-bridge-human.html

Convexity and symplectic geometry and the associated fields of theoretical
physics and mathematics play important roles in describing classical
mechanics, but also other areas of both physics and mathematics including
in information theory and optimization.
And symplectic packing is useful in optimization in confined volumes.

Which begs the question, if convex spaces point in a direction where finite
and infinite dimensional approximation, using either graphs or matrices are
used for data infrastructures upon which we unleash functions and
algorithms for machine learning supposedly emulating brain neuronal
architectures shouldn't we be looking at how data is processed and produces
information and all the looping operations which seem to create cognition
and even consciousness?

Thus the definition of knowledge and with it knowledge representation
spanning all current academic fields, each with their own paradigms for the
former two become a field of study of how bits and chunks of sensory input
are processed mathematically in convex space and symplectic geometry
settings.

Dave Raggett is definitely on the right path.

I have been feverishly working on notes to flesh out how we can do this,
and think it is fundamental,  because the concepts of knowledge and
knowledge representation need a firmer footing.

Milton Ponson
Rainbow Warriors Core Foundation
CIAMSD Institute-ICT4D Program
+2977459312
PO Box 1154, Oranjestad
Aruba, Dutch Caribbean