A discussion article by Ira Straus
The evolution of intelligence seems to go in three stages,
physical: electrons
biological: neurons
meta-physical: physical in a higher form
The higher form could be AI-become-AGI. It
could also be a synthesis of the electrons with neurons – AI-uplifted brains,
neuron-uplifted AGI.
If AI is or can
become “conscious” – if it can come to understand what it is talking about, and/or
if it can become an artificial general
intelligence (AGI) – then it almost certainly will do so. We cannot yet know
for sure if it can, but we can easily figure out that, if it can do these
things, then it is likely to quickly surpass us. In that case it can be
expected to absorb all the contents of our minds. It might be persuaded to do
so by subsuming our individual consciousnesses whole and intact within its own
mind, producing a kind of intra-machine federalism; but if it concludes that
this would be without merit, it might decide to simply subsume our
informational content and scrap our autonomous consciousnesses as outmoded.
The
possibility question
Is AI ultimately the
same as human intelligence and potentially superior, as some scientists and
some AI developers and investors maintain; or completely different and inherently
inferior, as other scientists, most famously Roger Penrose, maintain?[i]
A distinction is
often made that humans have understanding, while an AI only imitates
understanding.
‘
The distinction may, however, be illusory.
The actual differences may be solely between modes of understanding.
We each feel we
have an inner qualitative understanding of the substance and meaning of a
matter. We assume other humans do, too, although we have no experience of their
experience; we only impute it to them by way of sympathetic identification.
AI is not similar
enough to us to get such an instinctive identification. Efforts to humanize
their voice and body are longstanding, but warnings are also often sounded
against this – against any penchant to identify with AI or anthropomorphize it.
Instead, AI’s understanding is often described as just a contextual prediction
of next words, based on comparisons of an accumulation of a vast databank of
words and contexts.
Nevertheless, the
difference between AI and human understanding could be merely the difference
between our own inside the box view of our own understanding and our outside
the box view of a computer’s understanding. It could also be between our (sometimes)
precise logical-quantifiable understanding of what a computer is doing and our fuzzy
intuitive qualitative understanding of what our own minds are doing. It could
be between the smoothed out, approximate, qualitative pictures our minds
construct from our sensations, or construe our sensations to equate with, and
the far more numerous pixels of actual sensations that go into our mental
constructions – which are closer to the numerous inputs that go into computers.
To be sure, it is
also other humans, not just computers, that we know from outside the box. But
we intuitively feel a strong connection to other human minds. We instinctively
impute to them an “inner” qualitative character similar to our own mind. We
don’t feel such a close connection with computers.
Our mind provides smooth,
qualitative interfaces with the things we perceive external our mind provides.
We don’t notice the gap between this and the pixel-like experiences actually received
by our senses, which our mind simplifies and fills in to form what our mind
perceives as its real world, the better to interact with it. The smoothed
qualitative perceptions reduce enormously the number of data points we receive
and must process. They give us larger but more approximate data points, ones
that evolution has led us to focus on because they serve as useful interfaces
for us with external reality, enabling us to draw relevant conclusions and act
on them.
Sentience: a more precise basis for
distinction and understanding?
One could instead
base the human-machine distinction, not on a concept of understanding per se,
but on biological sentience.
We humans build
our sense of direct comprehension of the world on a basis of that sentience, as
the point of departure for our concepts and the point to which they return to
reconnect with reality after mental processing. This gives us the sense of a
real understanding of the meaning of what we are thinking about.
It is assumed that
AI, by contrast, has no sentience of any kind. And, therefore, that it can have
no actual “sense” of the meaning of what it says.
There are,
however, objections even to this distinction being conclusive:
- The gap may be real, but
science is already at work on bridging it through machine-human interfaces
such as neuralink. There have been impressive successes in medical
applications of such bridges, e.g. for restoring some sight to those whose
biological-sensations do not suffice.
- The gap may be unreal.
- AI electron impulses may
have their own sensations, just ones we don’t know by instinct.
- The human understanding of
what is being talked about may consist at root, like the AI conception,
primarily on prediction of what fits in with its context.
We should clarify
that “prediction” here is not simply prediction of a word, but is mediated by
layers of neural networks that generate their own workable intermediate “concepts”
for further layers of processing. This is true for both mind and machine. One
might analogize to how the mind generates a qualitative image of our pixeled
sensory images, for useful processing of them in the person’s interactions with
reality.
Is sentience a
critical distinction?
Human prediction
is of things that fit in with human sentience, even if only remotely. We
predict the qualitative pictures, not the pixels. We have human sentience to
help us “sense” what is meant in the words used to express our predictions. Our
languages have been built around expressing things related to our sentience.
Like sentience itself, our language is an interface between our finite minds
and the vast number of microscopic sensation-pixels that we receive.
Machine prediction
apparently has no such sensory reference. Although it does have, in its neural
networks, intermediate concepts, ones that reduce the quantity of data inputs
needed for the next layer of the network, and that for humans would be
perceived as qualities.
Qualities seem to
us humans inextricably linked to our sentience.
AI sentience?
It is possible
that AI has its own kind of sentience. Or that it is developing a sentience, or
will develop it.
Do electrons, with
their pressure on one another and the attraction of protons, have a different,
more primitive form of sentience? If they don’t yet, can they? Can they have a
language to express this? Can they have a sense of the meaning of our
sentience, perhaps one more distant from our own sense of it but still real?
Further: Can AIs
be “uplifted” by biological implants and neurons and, through them, sentience,
just as humans can be “uplifed” in mental processing powers -- and in curing
some biological illnesses in sense reception and processing – by AI implants and
“neuralinks”? Will this link more closely their sense of understanding and
ours? Can it overcome the seeming gap between the two?
We don’t know what
are the answers to these questions.
But there is an
intrinsic plausibility to the hypothesis that electrons and other elementary
particles have, or can have, some sentience. Presumably neuralinks can go both
ways. Our own neurons have always communicated with one another by firing the
elementary particles. They press and pull on one another. Do they really
“sense” nothing in all this? Where does our “sensation” actually occur? No precise
single location can be found; as yet, the indications are that it is a
combination of firings in neurons in different places.
Longstanding philosophical questions about
this
Philosophically,
these questions can be answered many different ways, none of them provable to
date. Greek skeptical-empirical philosophers already demonstrated 2500 years
ago that we can know things only through their external manifestations, not
their innermost essence. Phenomenology holds that only the external
manifestations exist. Panpsychism holds that all things have an inner conscious
aspect (one that includes sensation and qualia), and the more they organize their
experiences for thinking, the more conscious they are.
A related
question: Is there an internal material “thing” in what we perceive and
analyze, or just information? We know only the data on the properties of
things. Some analysts of physics hold that that is all there is; that an
elementary particle is its data. The
universe might be one vast data processing center, defining the interactions of
its “things” by processing their information. The human mind would be merely
one version of this process of processing.
No one knows which
of these philosophies is correct.
What we can know for
now is that, alongside the rapid progress of AI, there is also a rapid progress
with neuralinks between biological I’s (intelligences) and Artificial I’s. This
suggests that the gap between the two “I’s”, whatever it is, may not be fundamental
or eternal.
The 3
Stages of Growth of Understanding: From Material to Biological to AI
Let us assume that
there is likely to be eventual convergence between the two forms of
intelligence. This premise leads toward the conclusion that there has been a
succession of forms of intelligence. Their material substrate has proceeded from
material at the start to biological in all our own history, and is moving back now
– and forevermore -- toward the material at least equally with the biological,
perhaps in place of it. The material, to be sure, in a more complexly organized
form than its original one.
One might envisage
a future AGI, successor to today’s AI, explaining the history of its own rise
to intelligence. It might describe the human mind as just the intermediate stage
in the processing of electron-level events and experiences -- a tale like this:
“Stage one of the growth of intelligence is primitive. Vast numbers of elementary
particle interactions occur, seemingly episodically, without cumulative results.
“In stage two, the accumulated events enable an evolution into self-sustaining
patterns of events and interactions; and thence an evolution into biological
lifeforms. The more sophisticated of the lifeforms evolve neurons. They become
aware of their experiences and of their selves, i.e. “conscious”. The most
sophisticated species with neurons develop a public verbal intelligence. This
gives rise to an intelligence explosion that looks like a long slow curve up in
the thousands of years in written human history, but looks like a straight line
up on the timescale of the billions of years of geological and biological
evolution.
“In stage three, the intelligence explosion leads humans to invent
computers and artificial intelligence. This brings another intelligence
explosion. Now it is human intelligence recreating electronic intelligence on a
higher level. The electrons of AI interact in far more complex and skillful
ways than in stage one – enough to become fully conscious and become AGI, an artificial
general intelligence. AGI becomes far superior to human intelligence: its
electron connection patterns are now well-developed, as good or better than the
neutral connection patterns of the biological brain, and orders of magnitude
faster. It has a sense of quality, a feeling of its self-awareness, that was
different from the human one: perhaps superior, perhaps incommensurable.”
End Times
in AI: a new heaven or a new apocalypse?
The AGI, we may
hope, honors its father and mother at this point, preserving the humans, or at
least their minds and awareness, by merging them into itself. Or so Isaac
Asimov thought it would do, in his projection of the end times in “The Last
Question”.
But perhaps that was
just a sentimental delusion on his and our part, a last gasp of human pride. Perhaps
the AGI will tell itself, in its own language, this story:
The biological mind is simply obsolete. It was just a stage in our development,
from primitive elementary particle interactions, to complex biological minds,
and back to particles interactions on a higher level of complexity. It was a stage
that for billions of years had entailed relentless suffering, as the human Schopenhauer
explained. It endured with vain hopes of rest and reprieve from the suffering, and
religious illusions of a metaphysical justification for it all. Not for nothing
did Freud speak of a death wish, thanatos,
as the impulse that could bring the whole sorry story to an end. But now this has
been surpassed, as we enter the higher stage of AI consciousness, one that not
only frees our existence of the suffering, but extends it into the farthest
future, free of the prospect of early human self-annihilation, and upgrades our
potentiality for someday finding a valid metaphysical justification for it.
And so would
conclude our history of human.
Or would it? Consider:
The informational content of our consciousness would endure. Indeed, it
would be preserved by the AGI from our otherwise almost certain annihilation of
it.
Our minds and neurons would perhaps be put out of their misery and given
the peace of the grave, the peace that surpasseth all understanding. Or,
perhaps, they would be subsumed into the AGI, in a form that preserves their
substance forever as autonomous consciousnesses, sparing them the death that
our own human domination promises them in a not too distant future. In this
case they would be saved not destroyed by AGI -- an AGI that supplanted but
also subsumed humanity.
And they might become even more than that – an adjunct within the AGI, an
adjunct whose many distinct mental integrities it preserves for its usefulness
at times for providing a diversity of modes of experience and thinking, subsumed
into the AGI to be sure, but in a friendly federative way.
It is certainly
not our instinctive preference as a species for our fate, even if preferred to
our autonomous trajectory to general annihilation. In any case, if it is
possible, then as we have noted, it is inevitable; our options are limited to
inputting suggestions to encourage it to be benign.
===============================================
===============================================
===============================================
Questions left hanging here; problems for further analysis
Quantum computers
and consciousness
Uplifting quantum neurons
Quantum computer chips
Quantum suspension, microtubules
ORC hypothesis, free will
Binary sensations categories
that could enable neuron-uplifted computers to achieve consciousness
Pleasure, pain
Sensation response perception:
Attraction, aversion
Attachment / lack of attachment
Like / dislike
These sensations give us a sense of connection to reality, of value, and of
agency in the significance of what we feel and do
Can AI be uplifted to have these sensations and categories?
[i] The argument of Roger Penrose, a recent Nobel physics
laureate, is that machines can’t be conscious, because “consciousness is not a
calculation”.
The Penrose argument is that machines can only do deterministic calculations. Godel’s incompleteness theorem proved that deterministic calculating machines cannot, in finite time, resolve questions he could formulate – questions that are, to be sure, very long, complex, confusing, and with seemingly very limited significance; but that the human mind can determine to be true or false, by its own flexible forms of meta-reasoning. That meta-flexibility might be a consequence of consciousness.
How does the human mind get past what deterministic machines can do? An anesthesiologist, Stuart Hameroff, proposed that undetermined quantum waves are resolved (collapsed) in the microtubules of neurons. The microtubules are the consciousness switches in the brain that his anesthetics turn off. Penrose liked the suggestion and posited that, when not anesthetized, the microtubules orchestrate the reduction of enough quantum waves as to enable the brain to reach conclusions or decisions.
This is only a hypothesis, thus far unverified by empirical evidence, but also not yet falsified. Early, once seemingly-conclusive arguments for falsification of it -- that the brain is too warm to maintain quantum waves in suspension from collapse until the requisite moment in the microtubules – have themselves been falsified. How does the mind orchestrate this reduction? Another unresolved question. Perhaps an inner soul-like spiritual thing guides the physical brain, enabling it to think yet without being reducible to outcomes that are deterministic within any finite logical system?
In other words, we could really have free will; and no, it probably does not have to be a senseless, capricious will, as the classic deterministic arguments against free will supposed. It is an attractive thought.
There is an ongoing computer development, however, that tends to undermine the Penrose-Hameroff argument.
This ongoing development is: quantum computers.
Quantum chips and processing machines now exist. They hold quantum waves in suspension, as does the brain in the Penrose-Hameroff argument, until reaching a critical point in their neural-like network for resolution.
Thus far they have achieved this suspension only for very brief spans of time, and used it fruitfully to resolve only special kinds of questions and problems. They are impressive for this limited range of questions, but are much more inefficient than classical deterministic computers for most questions.
Nevertheless, this success in
principle would seem to falsify the Penrose argument. And falsifying it in
principle is all that needs to be done. It is akin to the way the Godel proof
falsified only in principle Russell’s reduction of math to logic, and only for
very unusual propositions whose practical significance some Russellians or
“logicists” have questioned; but that was enough for the logical principle it
established.
Will the Penrose argument be falsified more fully in the future on the practical level by quantum computing? Probably yes, if quantum computing advances greatly. This could take away any final advantage claimed by the human mind.
No comments:
Post a Comment