Ok, lets try to assemble a bunch of random thoughts in a somewhat
coherent manner and call it an post...
The current state of affairs is roughly as follows:
Right now we have proto-AGI systems that are showing some promise.
However, they are being trained on exaflops scale supercomputers at a
cost of about a million dollars for the training. The result of this
training is somewhat like a moderately coherent 90 year old who is able
to tell tales of the glory days of 2021 but can't remember what he had
for breakfast...
In my previous message I proposed turning things around and using the
input stream as the training feedback set, allowing the thing to run in
a continous unsupervised training mode. I am fairly sure this is how the
brain works and would be a candidate for consciousness.
The distressing thing tho is that the current compute requirements mean
that you need to be able to dump many millions of dollars into it. Now
we can probably claw a good 20% of that back with algorithmic
improvments and low level coding tweaks but still that's a hell of a big
barrier. An AGI will always be in training mode so there is no cheap
"inference mode". It always has to be training, but on the flip side
there is no massive pre-training period, it's basically learning on the
job from teh start like humans do. We're still talking about an exaflops
capable machine though.
I do not have a cellular phone and that is a requirement, for some
reason, for trying out *-GPT. =\
As a black-box AI, it is pretty broken in many ways but also quite
superhuman in others.
We now need to turn our attention to exactly what we mean by
superintelligent. There are at least two distinct classes of
superintelligence.
for starters consider an AGI that is legit verified human equivalent, at
least within a factor of 0.8 on all dimensions but is 2x of the smartest
known human baseline in one of those dimensions. Well that is a legit
superintelligence but I'd call it a weak superintelligence.
A strong superintelligence will have capabilities on dimensions that do
not exist in the human baseline.
These capabilities include the ability to compute using more powerful
algorithms than the human brain can. These capabilities also include the
ability to operate multiple concurrent streams of thought, having proper
distributed intelligence, and the ability to use modalities that are not
available to the human baseline, even with neural interfacing.
Ok, now we're going to have to hop on the magical airship and take a
trip into philosophy land. Philosophy land is a tricky place with lots
of religious and emotion-driven ideas. Lots of it is not actually
defensibile but are fiercely protected anyway. Of the ideas that are
logically defensible, there are still value judgements that can have
profound consequences.
A criticism of my thought would go along the lines of pointing to my
focus on subjective consciousness as trying to protect a fiction at
great cost, typically from brain uploaders. My criticism of uploaders is
that uploading is a fatal error that cannot possibly preserve one's
subjectivity and that the clone produced by the procedure would not
enjoy enough benefits to make the loss worth further contemplation.
That said, once I get an AGI/ASI working well enough is to start
seriously working on what I call the "arcanum of consciousness". In my
fevered imaginings I picture it as a round stone tablet engraved with
runes. It would basically function as a rosetta stone and state-map of
consciousness. It would be able to decide whether a thing is conscious
or not, and it would specify how that consciousness can be reconfigured
and what changes are prohibited. While it can't contain every possible
answer, it would specify parameters that can be fed into something like
a propositional logic engine and if a solution can be found, then the
mind in question can be evolved in the specified direction.
That, finally brings us to human enhancement. The Elon is absolutely
correct with the concept behind neuralink. The problem is that there are
a fairly wide variety of brain types out there. There are a few neural
typical individuals in the world but there are many many many different
brain-types out there. Temple Grandin (sp?) is a famous example but
there is a fairly wide diversity of reports about what conscious
experience feels like.
I see the Arcanum as a solution to this problem, it would be able to
classify brain types, lets imagine that categories A B C and D are
identified with minor types e, f, g and h. Furthermore individuals may
have a specific brain type that works like an A-brain but one or two
pathways are either "dominant" (ie expressing increased activity and
conscious focus) or "suppressed" (ie expressing little or no activity).
Ok, given an individual with an identified suppressed pathway, what
would happen if that pathway is artificially repaired or stimulated?
Ok, proposal on the table is to take this ASI framework and connect it
to a given patient. Lets say that the patient has a type B brain but
neural circuit 3 is suppressed or absent and circuit 2 and 4 are
compensating so the patient can operate within human norms (or not, as
the case may be). I see the Arcanum as a tool to map out what
connections and enhancements must be made to properly connect this
unique brain to the ASI enhancement.
Without the arcanum, then you are basically limited to connecting to the
peripheral nerves, and cannot provide the patient a menaningful
(worthwhile) enhancement to his cognition and that the ASI is limited to
assisting the patient as a social agent, which feels far less satisfying
to me.
Anyway, this post has probably long since passed into TL;DR category....
Where's the send button?
--
Beware of Zombies. =O
#EggCrisis #BlackWinter
White is the new Kulak.
Powers are not rights.
------------------------------------------
Artificial General Intelligence List: AGI
Permalink:
https://agi.topicbox.com/groups/agi/Td72675df603b4d93-M46d5ac3e6b524681cee2ec90
Delivery options: https://agi.topicbox.com/groups/agi/subscription
