On Sat, Jan 17, 2015 at 6:37 PM, Tim Tyler via AGI <[email protected]> wrote:
>> de Gray is arguing against the scenario where a recursively self
>> improving AI in a box goes FOOM!
>
> It seems like a straw man scenario. Has anyone seriously proposed it?

There are some old proposals, for example Corwin's experiments on
containing AI in 2002.
http://www.sl4.org/archive/0207/4935.html

Yudkowsky's Coherent Extrapolated Volition in 2004.
https://intelligence.org/files/CEV.pdf

Of course this was before Google and Facebook got good at recognizing
images and natural language text. It was not so apparent then as now
that the internet is becoming AGI. Of course you cannot contain it or
turn it off. In order for AGI to gain human knowledge, it has to
interact with humans.

Also, the internet looks like much less of a threat than a paperclip
maximizer. We understand that AGI is not a powerful optimization
process with a simple goal. We don't tell computers what to do. We
tell them how to do it because it is faster to share our own knowledge
than for it to figure it out on its own. There is simply no such thing
as a general purpose learning algorithm that you can give an arbitrary
goal to. Mathematically, there never will be.

>> Most of what AI in general already knows comes from humans. AI cannot
>> learn human knowledge faster than humans can communicate, about 5-10
>> bits per second, but that is faster than anything else.
>
> Machines can easily learn about all the images and videos freely available on 
> the
> internet. They can slurp that information up over as many T1 lines as you have
> going into your data center.  There's no 5-10 bits per second limit.

Machines can only learn rapidly the things we already know. There are
some important questions that even very powerful computers cannot
answer quickly. Among the most important are how we can live longer.
Ray Kurzweil takes 100 pills a day, hoping to live to see the
singularity in 2045 and be immortal. But there is a problem. There is
not a single pill of any kind that is known to increase life
expectancy. It would take decades to find out.

Sure, we have in-vitro and animal models. We can learn very quickly
that calorie restriction extends the life spans of fruit flies and
mice. We don't know if it works on monkeys. After decades of
experiments, it worked on one group but not another. There have been
no experiments on humans. We know that children reach puberty at a
younger age now than 100 years ago, probably due to more calories, but
people are also living longer. Calorie restriction probably works by
slowing growth. Fish convert 90% of what they eat into growth and 10%
to energy. Cattle convert 15% of their food to growth. Humans convert
0.3%.

We suspect that rapamycin and sirtuins mimic calorie restriction, but
the results are not conclusive and these drugs can have serious side
effects. Rapamycin suppresses the immune system. Some studies say that
light drinking increases life expectancy. Others say that any alcohol
increases the risk of cancer. We once thought that vitamin supplements
helped, but later studies proved that false. We thought that low fat
diets helped, but later studies found they did more harm than good. We
thought that a low salt diet helped, but later studies refuted that.
We thought that sunscreens helped, but skin cancer rates have been
increasing in tandem with sunscreen use. It turns out that sunscreens
block UVB but not UVA, and actually increase exposure to UVA (95-99%
of UV depending on angle of sun) because it is UVB that gives you a
tan, and a tan blocks UVA. SPF is rated by UVB blockage only.

My point is that all of this knowledge took decades to learn, and the
problem is getting worse. New drugs now cost $2 billion to develop.
The cost doubles every 9 years. The rate of increase of life
expectancy has peaked at 0.2 years per year in the 1990's and is
declining. It peaked in the 1970's in developed countries.

One would hope that we could build computer models of the human body
that would allow us to answer these questions faster. But we do not
have computer models of even simple chemistry. There is no program
that inputs a formula like H2O and calculates the freezing point of
water. The reason is that modeling the movement of atoms requires
solving Schrodinger's equation, which has exponential time complexity
except on a quantum computer. But even a quantum computer is no faster
than doing the actual experiment because all you have is an
exponential speedup of an exponentially slowed down algorithm. A
simulation always requires more computation and is less accurate than
the actual experiment.

And of course, the human body is much more complex than H2O. The brain
executes 10^16 synapse operations per second on 10^14 synapses. The
body executes 10^20 DNA, RNA, and amino acid operations per second on
10^23 bits of DNA. Even if Kurzweil's prediction of computing capacity
catching up with the brain in 2045, we will still be a long way from
simulating the body. After robots automate everything that humans can
do with their brains, senses, and muscles, there will still be plenty
of jobs testing experimental drugs.

-- 
-- Matt Mahoney, [email protected]


-------------------------------------------
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