om

Tonite me wanna mumble about the issue of hardware overhang that we're
now facing.

One long-standing worry in AGI research is that if the date of it's
development is delayed too long then, once it is developed, that it
would quickly become vastly superhuman before it could be
studied/debugged/understood/controlled, etc.

Have a look at:    http://www.cryengine.com/

This is what our current hardware can do, if it weren't being programmed
by the chimps who make desktop applications. A top-end graphics card can
crank that out in ultra-high resolution at up to 60fps...

So, who among you is going to stand up and state a serious case that
what we have now in a high-end desktop or a $15k compute-optimized
server isn't anywhere close to achieving human intelligence?

One issue is the uploader's mentality. They derive their estimates for
general AI from their estimates for brain emulation. As I've said many
times before, and here again, A pure AGI architecture will be several
orders of magnitude more efficient than any conceivable brain emulation,
except those that have been abstracted so far that there is barely even
the most tenuous resemblance to the original. Superhuman AI will be
available at least a decade before mind uploading and mind uploading
cannot result in a superhuman mind (without violating the premise on
which the argument in favor of identity is based).

The reason brain emulation is so expensive is that it is trying to
emulate synaptic junctions when the actual unit of functionality is the
neural ensemble (on the order of a few hundred neurons). It is difficult
to quantify the information content of neural signalling but the key is
that it takes dozens of neurons to communicate just a single scalar
value. The brain does things that way because of the uncertainties of
the metabolic environment and the necessity of maintaining extremely
high reliability over the lifetime of the individual, irrespective of
the life-cycles of the individual cells.

Modern computers operate on very different principles and therefore can
only be compared to neural circuits in the broadest outlines.

Just about every area of the brain that has been reverse engineered to
the point where we can say "this is a circuit for doing X", we have
already surpassed the measured performance by many orders of magnitude.

The only reason that this hardware overhang hasn't been recognized is
that people still suck at programming and because people are very
protective of their egos. If they felt inferior to a pile-O-parts, that
would make them feel really bad. So therefore they point at every
available example of AI failing and pronounce that it is because our
biology is that much awesomer. -- It's not. Even now.

Okay, now let me hit you with a new, and much more important, concept. I
call this concept "Algorithmic Backlog" For hundreds of years now, we
have been developing mathematics and algorithms that, in many cases, are
much more effective at solving a wide variety of problems than the
general purpose but, ultimately, approximate pattern matching techniques
our brains use. The most important qualitative difference between AI
thinking and human (or upload) thinking is that the AGI can incorporate
many of the algorithmic advances we have made over the years directly.
It's not clear exactly how many such algorithms can be applied but It's
a pretty sure bet that the answer is more than none.

It is not at all clear what the overall implications of all of these
issues are, except for one thing. AI can happen Real Soon Now.

-- 
IQ is a measure of how stupid you feel.

Powers are not rights.



-------------------------------------------
AGI
Archives: https://www.listbox.com/member/archive/303/=now
RSS Feed: https://www.listbox.com/member/archive/rss/303/21088071-f452e424
Modify Your Subscription: 
https://www.listbox.com/member/?member_id=21088071&id_secret=21088071-58d57657
Powered by Listbox: http://www.listbox.com

Reply via email to