Re: [sl4] FAI development within academia.

From: Roko Mijic (rmijic@googlemail.com)
Date: Sun Mar 01 2009 - 09:36:07 MST


2009/3/1 Roko Mijic <rmijic@googlemail.com>:
> 2009/3/1 Matt Mahoney <matmahoney@yahoo.com>:
>>
>> --- On Sat, 2/28/09, Roko Mijic <rmijic@googlemail.com> wrote:
>>
>>> Today there are about 100 people in the world working
>>> on AGI in a vaguely serious way, as far as I can tell,
>>
>> That's if you don't count Google, which is slowly developing AGI but not calling it that, and the semiconductor industry, which is building the computers needed to implement AGI, and the billions of people posting information to the internet, which is providing the knowledge base that AGI will ultimately use to make it work.
>>
>> In case you haven't noticed, the internet is gradually getting smarter. We have trouble defining what "smarter" means, but as long as we insist on defining it as "more like a human" then we will be left behind by a singularity seemingly without ever achieving AGI.
>>
>
> I'm intrigued by this point of view, though I currently don't place
> much credence upon it.
>
> Can you make precise what you mean by "the Internet is getting
> smarter, ... we have trouble defining what smarter means"
>  - presumably if you don't know what smarter means, then you can't
> claim that the Internet is getting it...
>
> I take my definition of intelligence to be "the ability to achieve
> goals in environments". According to this definition, the Internet has
> no intelligence, because it is not an agent and hence cannot achieve
> any goals in any environments.
>
> It is the case that the WWW contains a lot of knowledge, such as
> wikipedia, google books, etc. But knowledge alone is not intelligence
> - would you consider a very big library to be intelligent? Would it
> worry you that adding more and more books to a large library might
> cause the Singularity?

So, perhaps the strongest case I can make for the proposition I think
that you are putting forward is that once the internet accrues
sufficient knowledge, it will be very easy for someone to write a ML
based AGI using said data.

But even this I doubt. Yes, you might write a ML based system that has
a very good prior over, say, human semantic knowledge, (i.e. a very
good statistical model of our world) but this would not be a finished
AGI. It would not know how to form plans, to excecute actions, to
learn new things, to debug its model of the world, to debug its own
behavior or reflect on its internal processes.

>
> I realize I could be wrong here.
>
>
>> -- Matt Mahoney, matmahoney@yahoo.com
>>
>>
>>
>
>
>
> --
> Roko Mijic
>
> MSc by Research
>
> University of Edinburgh
>

-- 
Roko Mijic
MSc by Research
University of Edinburgh


This archive was generated by hypermail 2.1.5 : Wed Jul 17 2013 - 04:01:04 MDT