[ExI] Kelly's future

[Samantha Atkins]

On 05/26/2011 01:49 PM, Kelly Anderson wrote:
> On Mon, May 23, 2011 at 3:47 PM, Damien Sullivan
> <phoenix at ugcs.caltech.edu>  wrote:
>> On Mon, May 23, 2011 at 03:22:02PM -0600, Kelly Anderson wrote:
>>
>>> Once achieved, an AGI is easily replicated. That much I will grant
>>> you. But mixing explicit programming with a training process is very
>>> difficult. Just look at how hard it is to change people. Changing me
>>> from a religious zealot to an atheist was a very painful process that
>>> took a couple of years of hard work. It was not just "changing the
>>> programming", although that was, in a sense, exactly what it was.
>> While the results of machine learning may well not be easily modifiable
>> or reverse-engineerable, using people as evidence isn't very good.  We
>> don't have explicit programming of people, or of brains, the way we do
>> have of computers.  Verbal instruction is a limited ability, compared to
>> being able to go in and change neural wiring directly.  Not that we'd
>> know much what to do if we could, but we don't even have the safe access
>> for people.  Whereas even a genetically evolved neural network mess of
>> code is completely open to our examination and modification.  Knowing
>> what to do is another matter, but the fact that we can't do things to
>> people through their skulls is kind of irrelevant.
> Yes, what we can do to people is irrelevant.
>
> My underlying assumption here is that AGIs of the future will
> (initially at least) be based upon the human intelligence model, and
> will, in fact, be silicon based human brain emulators.

I think this is much much harder to do than a more direct approach to 
the parts of intelligence we are after.  It seems as doubtful as 
achieving heavier than air flight by building a bird emulator.

One problem in a brain emulator that is pretty critical is how to 
separate all those things laid down by evolution that you do not want, 
or at least want in a more optimal form, from those you do want.  One 
problem is that currently it would take a 4GW power plant just to run a 
cat brain partial emulation.  We need some device breakthroughs like 
perhaps memristors.

- samantha