On Fri, Feb 28, 2014 at 11:06 AM, Ben Goertzel <[email protected]> wrote:
> See
>
> http://hplusmagazine.com/2014/02/28/saving-the-world-with-analytical-philosophy/

And my blog comment. I read the book and reviewed it on Amazon.
http://www.amazon.com/review/R1BC9TNLJSULFU/ref=cm_cr_rdp_perm?ie=UTF8&ASIN=B00IB4N4KU&linkCode=&nodeID=&tag=

I still find it annoying that MIRI models AI as a powerful
goal-directed optimization process, when in fact this has nothing to
do with how AI is developed in practice. We don't build general
purpose reinforcement learners and let them figure out how to satisfy
some complex goal like winning at chess or Jeopardy. It makes no sense
from an information theoretic point of view to transmit a large number
of bits from the human brain to a goal system through slow channels
like speech or writing, and then transmit that information one bit at
a time to an AI through a reward signal. The approach we use in
practice is to update the AI through explicit instruction to meet the
goals in the developer's mind. The AI learns faster when it doesn't
have a goal.

This is a problem of information theory. It won't be solved by
recursive self improvement because an AI rewriting its code adds no
bits at all. So MIRI needs to explain why this will change in the
future if it believes that unfriendly AI is a threat. If faster
hardware will make reinforcement learning practical, then keep in mind
that human evolution required 10^48 DNA base copy operations, which is
higher by a factor of 10^22 than the number of equivalent neural
network operations performed by a human brain over a lifetime. If the
threat is from self-replicating agents (which I think is real), then
the book should mention it.

-- 
-- 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