"So exactly what do you hope to accomplish?"
Great question Matt. It haunts me and I imagine that others are
considering what their answer might be.
In re-reading your post there are a couple of statements that help to
clarify for me what I want to accomplish. For example:
"Automating labor and extending life through uploading both require
models of the human mind, which includes stuff like vision, language,
robotics, and art."
This statement doesn't go in the direction I want to go. I don't see a
"model" of the human mind as a goal. First, the understanding of the
human mind is pretty slow in coming. Rather than base artificial
intelligence on the design of the human mind, why not base it on the
"task" of being intelligent? The human mind may be a good model
eventually, it surely has better compute energy economics than anything
we are doing.
I suspect that as we design components that can perform the intelligence
task we will gain insight on what the mind is doing. Once we get the
clue, then science will find clever ways of confirming how the mind does
it.
Another statement:
If you or your loved ones decides to upload you
after you die, then a model of your mind becomes a component of a
program for a robot that carries out its predictions of your actions.
If you think about the implementation of this, there are a few
problems. One that strikes me is the instability of most human minds.
We talk of predicting what our actions would be, but many times we don't
know what we will do in a given situation and what we do may be for
reasons that are very subtle and unexpected. How many times do we do
things that we didn't expect to do?
Rather than build a model of me that has to be endlessly tweaked to get
my behavior, I would approach the upload task (I don't care about
uploading, God can handle the life after death stuff) as a project to
know what "me" sees as opportunity in my day to day living. This brings
up an issue, does anyone want to know what I would do?
Does anyone care about watching me act my role in life? Not likely.
But, would anyone want to watch closely and see what "The Most
Intelligent Known Entity" would do with itself in day to day life? MIKE
sounds pretty scary, but I suspect MIKE wouldn't be that much more
aggressive than the Bill Gates and Mark Zuckerbergs of the world. Not
that Bill and Mark are so smart, but they have their money and freedom
to do what they think is best. So far they don't think it's best to
conquer the world of human affairs.
The point is, we don't care that much about Bill and Mark's thoughts and
directions because we consider them as not that much different than us.
What confuses and confounds us, probably confuses and confounds them.
But if we had a unit that was producing intelligent behavior, and could
be instrumented to reveal motivating factors, that would be interesting.
What I hope to accomplish: dabble in AGI and make a good case that we
can build intelligence because we have an architecture that produces
behavior we would all agree to be "intelligence" in action. The
architecture isn't about building a Sim City kind of mind which models
the world (admittedly I don't know much about current Sim City
programming.) The architecture is about collecting these bits of
knowledge of what is important (opportunity) and using that knowledge to
navigate through the world - be it a world of networks, robotics, or
human affairs.
No wonder no one else is venturing into what they hope to accomplish...
.02.
Stan
On 11/1/18 11:31 AM, Matt Mahoney via AGI wrote:
Abstraction is not simple. AGI is not simple. I realize that most of
the interest in AGI on this list is out of scientific curiosity. But
AGI is expensive so anyone funding your research is going to be
looking at practical applications like automating human labor or
extending life. We are already making progress on both fronts, which
means that the easy parts have already been solved. Automating labor
has a ROI of world GDP divided by interest rates, or $1 quadrillion.
Extending life has a ROI of world GDP times life expectancy, or $5
quadrillion. Google, Amazon, Apple, and Facebook have a combined
market cap of $3 trillion, so they have a long way to go.
Automating labor and extending life through uploading both require
models of the human mind, which includes stuff like vision, language,
robotics, and art. Basically it is a function that takes sensory input
and returns a prediction of what a human would do. It is required for
effective communication between AGI and its human masters so that it
knows what you want. If you or your loved ones decides to upload you
after you die, then a model of your mind becomes a component of a
program for a robot that carries out its predictions of your actions.
The rest is engineering. We already know that the best vision and
language models are neural. A human brain sized neural network with
10^14 synapses and 10 ms response times requires 10 petaflops and 1
petabyte. Computers with this capacity require 1 MW of electricity
costing $100 per hour, which is not competitive with human labor. A
human body averages 100 watts (of which 20% powers the brain). You
cannot get there with transistors because they are already not much
larger than atoms. Computing AGI will require moving atoms or ions,
not electrons. Is anyone working on that?
No. Everyone is working on the software, as if they expect it to run
on their laptop. Sorry. Even if you did have appropriate hardware,
humans are complex. Your DNA compresses to the size of 300 million
lines of code. If there was a simple formula for intelligence, we
would have evolved a lot faster or we would see it in insects and
bacteria. Legg proved that powerful predictors are necessarily
complex. Every example of AI supports that proof with more evidence
that you aren't going to find a simple equation for intelligence.
Watson was a 30 person-year effort. Google has written 2 billion lines
of code. Human evolution took 10^48 DNA base copy operations over 3
billion years.
So exactly what do you hope to accomplish?
-- Matt Mahoney, [email protected]
------------------------------------------
Artificial General Intelligence List: AGI
Permalink:
https://agi.topicbox.com/groups/agi/T586df509299da774-M4ca93c7e9f6ecc2547387763
Delivery options: https://agi.topicbox.com/groups/agi/subscription
