[Fis] Google DeepMind wants to define what counts as artificial general intelligence

[Marcus Abundis]

Dear Pedro,

Thanks for this information, I read the paper with great interest!
In my 24 October post (Defining Intelligence . . .) I noted I am developing
a related paper, which I have since provisionally titled 'Entropy — A
Simplified Scientific Base for Super-Intelligence'. I copy the DRAFT
'Introduction' below for those interested.

I believe there are innate problems to the approach posed by the authors of
the paper you noted. Reasons for why I think their approach is problematic
is covered (in brief) in the copy below.

===
ENTROPY— A Simplified Scientific Base for Super-Intelligence
by Marcus Abundis, Bön Informatics
DRAFT Paper — ver. 28Nov23
(May? 2024 NeurIPS DEADLINE: 10pt, 8 pages max, + 1 reference page)

ABSTRACT: This paper frames a top-down science-based approach to
Super-Intelligence, versus more-typical complex/fragmented anthropic and
statistical bottom-up methods.

INTRODUCTION — base issues, key terms, and central goal/method
Pondering the advent of Super-Intelligence (SI) holds many issues. First,
defining human intelligence (HI) is already daunting with many roles seen
in diverse individuals and cultures across the globe—often ignoring other
`intelligences'. Second, a core from which SI arises must be named—with
`general intelligence' (GI) as a likely prelude. But if global HI is so
elusive, how do we hope to define `cosmic GI'? Third, likely SI risks must
be noted. These are a few issues raised by exploring SI. Nick Bostrom
defines SI as: `any intellect that greatly exceeds the cognitive
performance of humans in virtually all domains’, which does not clarify
these issues but poses a base proposition. It omits needed detail on SI's
likely advent, but which this study targets.

This paper names a scientific base for general and super intelligence, that
also frames related risk and challenges. But as many mixed SI,GI, HI, and
base `intelligence’ views already exist, I first define my terms. These
initial terms are further expanded on over the paper’s course.

Key Terms:
General Intelligence (GI) — is knowledge of how things generally work and
fall-apart, `material functioning’ in the cosmos, sans `logical gaps’. The
cosmos and GI hold myriad direct contiguous events. But science holds
narrow measurable-and-repeatable roles, omitting `uncontrolled variables’
for repeatable and verifiable results. GI is thus the ideal science pursues
as pure Natural Philosophy. Bit for now, GI marks nearly-fanciful perfect
knowledge of the cosmos—essentially, Kant’s `das Ding an sich’.

Super-Intelligence (SI) — is knowledge of how things might work and might
fall-apart as `creative functioning'. Knowledge of creativity shows first
as partial GI—that SI expands via `latent functions’, exploring GI roles.
For example, one may imagine the Sun swelling to engulf the Earth as a
future event, or see birds in flight trans-scribed as a 747 jumbo jet. SI
thus surpasses manifest material reality, toward future (often human)
material possibilities. Traditional `science’ offers no such formal
creative narrative.

Knowledge — is `a grasp' of direct cosmic events, by indirect `referential’
means. References are: 1) direct events held in an abstract informatic
role, 2) often processed toward targeted effects, 3) `tried out' in
evolutionary environs: `agent’ stimulus-process-response (SPR). Non-agent
particles, atoms, etc. are energy-matter directly driving environs, that
agents survive. Ideally, all references (genomic code, mind as memory,
etc.) are jointly processed. For example, genomic shifts may yield `longer
legs’, but one must instinctually/willfully use `new legs’, for new
effects/knowledge. In turn, that joint work frames an agent’s sensorium and
ensuing `habitats’—all marking Kantian `bounded phenomena’ as the root of
all knowledge.

Human Intelligence (HI) — mixes instinct, thought, myth, and fact, with
creative-to-dull and solitary-to-social traits, alongside GI/SI clues. This
mix makes HI hard to typify and inapt for clear GI/SI modeling. But it also
implies adaptive plasticity and robustness, with agents regularly evincing
partial GI as `survival’ (via genomic and mind references). Here,
functional effectiveness-and-efficiency of references sets `one’s habitat’.
But also, extending one’s references (via genomic, mind, or SI `tools’)
expands one’s habitat— all of which marks a so-called Anthropocene and
typifies HI.

The means by which we extend our sensorium (via instinct, thinking, myth,
fact, etc.) is a fascinating topic, but ultimately unhelpful in framing GI
or SI. HI is thus seldom referenced in this study, departing sharply from
other AI approaches, most of which are anthropic in character.

Central Goal and Method:
In sum, the above means better knowledge/reference maps come closer to true
GI and SI. Also, all agents hold `some GI’, with extinction the only
alternative. Hence, differences in `how’ and `how many’ references one
maps-and-uses drive agent `intelligence’ (knowledge, meaning, etc.)

Base `intelligence’ meets most regular needs, but we also map creativity.
An eternally dynamic cosmos has extinction as an evident rule, with agent
survival the exception. Cosmic chaos/complexity make `better creative maps’
of also vital to agents extending survival. But mapping creativity is
notably problematic. Hence, a full map of regular-and-creative (GI and SI)
functioning, in the cosmos and in agents, is this paper’s central goal.

>From the above sense of ‘how things generally work and fall-apart’, two key
scientific roles mark likely GI and SI `informatic processes’. On how
things work, Claude Shannon’s Signal Entropy is often called ‘the mother of
all models’, to suggest one likely process. On things falling-apart, Ludwig
Boltzmann’s thermodynamic entropy is widely agreed and has high intuitive
appeal, and thus suggests a second likely informatic process. As such,
general knowledge of how things work and fall-apart, and creativity amid
the two, frame most aspects of intelligence and Nature— to frame this
paper’s central claim.
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://listas.unizar.es/pipermail/fis/attachments/20231128/508b49bb/attachment-0001.html>