On Thu, Aug 30, 2012 at 4:53 PM, Anastasios Tsiolakidis
<[email protected]> wrote:
> 1) which breakthrough/milestone you think is/are needed on your way to AGI

As I explained in my design, there are 3 things that must happen
before machines will be able to do all the work that we now have to
pay people to do because machines aren't smart enough.

1. A 1000 to 10,000-fold decrease in the cost of computation, memory,
and bandwidth. (I estimated 1 million in 2009).
2. Solutions to hard problems in language, vision, art, robotics, and
modeling human behavior.
3. A global system of public surveillance to collect human knowledge.

The estimated cost is $100 trillion to $1 quadrillion, mostly for
knowledge collection, once hardware prices drop. We need about 10^26
OPS and 10^25 bits of memory to run the equivalent of 10^10 human
brain sized neural networks, barring the discovery of more efficient
solutions to step 2. (So far none are known). Currently that much
hardware would cost about $1 quintillion, but should drop to $1
quadrillion in 15 years at the current rate of Moore's law. That is
near the break even point for investors (world GDP divided by market
interest rates).

The software (the focus of most AGI efforts) is the least expensive
part. A human genome with 3 x 10^9 base pairs is equivalent in
complexity to 10M lines of code, which would cost about $1 billion. To
model variability in the human population, we need the following
additional information per person:

1. Your parents: 33 bits each.
2. Which parent you inherited each chromosome from: 23 bits.
3. Mutations. Mitosis has a copy error rate of 10^-9 per base pair, or
3 mutations per haploid. It takes 32 bits to describe the location of
a SMP and 2 bits to describe the substitution. Crossover, copies, and
deletions require a similar number of bits. It takes about 40
generations for a man to produce enough sperm over a lifetime, or 4000
bit errors inherited from the father. The number of eggs is much
smaller, so I won't count these.

Thus, it takes about 10 times as many bits to encode the genome of
every person on earth as to encode 1 person. Thus, the software should
cost $10 billion, still a trivial share. No doubt it will be higher as
we first try to optimize for insufficient hardware, and then later for
insufficient human knowledge as this becomes the bottleneck. A rough
estimate of this cost:

1. 10^9 bits of long term memory per human (word, picture, and sound
recall tests by Landauer).
2. 99% of human knowledge is shared (U.S. Labor Dept: replacing an
employee costs 1% of lifetime earnings).
3. Speech, writing and typing transmit 2 to 5 bits per second.
4. Global per capita income is US $5 per hour.
5. Human population is 7 billion.

Thus, capturing the 10^17 bits that make each person unique would cost
about $100 trillion. This cost will go up as the economy grows and
time becomes more valuable. I think this will be the most severe
bottleneck. If we automate 90% of human labor, then the 10% that
remains might be 10 times as valuable, keeping the cost at $100
trillion for the remainder. This is why we need technology such as
advanced, high resolution surveillance that recognizes faces and
voices. Alternatives such as brain scanning would currently have to
cost $10K to be competitive, but might still be a long term solution.

> 2) which module(s)/subsystem you are certain will make it all the way
> to AGI 1.0, so which current theory/tech is essential to your vision

I think most of the theoretical work has already been done. Once we
have sufficient computing power, hard problems like vision and
robotics won't be so hard anymore. The hardware is a physics and
engineering problem. Clock speeds have already stalled, and transistor
sizes are now only tens of atoms across. Power and heat dissipation
are major problems. The world's most powerful computers are able to
run a human brain sized neural network (10 petaflops, 1 petabyte), but
need 10 MW of power, compared to 20 watts for the brain.


-- 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-c97d2393
Modify Your Subscription: 
https://www.listbox.com/member/?member_id=21088071&id_secret=21088071-2484a968
Powered by Listbox: http://www.listbox.com

Reply via email to