From: Nick Tarleton (email@example.com)
Date: Tue May 27 2008 - 20:20:35 MDT
On Tue, May 27, 2008 at 9:06 PM, Matt Mahoney <firstname.lastname@example.org> wrote:
> --- Michael Anissimov <email@example.com> wrote:
>> There's no way we can estimate which input rates correspond to which
>> cognitive enrichment rates.
> Yes you can, if you measure intelligence by how much you know (in bits).
> The intelligence of a child AI cannot exceed the intelligence of the
> parent plus accumulated input.
And why on Earth would you define it that way? That's knowledge -
intelligence is how well that knowledge can be acquired and used.
> I realize there are other definitions of intelligence. You could upload
> to a computer, then raise the clock speed, add memory, coprocessors,
> faster I/O, and a faster network connection. You could then think and
> learn faster, but initially you would not know any more than you do now.
Also, create better algorithms. See
http://www.sl4.org/archive/0606/15277.html . That may be nitpicking
here, though, since no-free-lunch theorems might bound increase in
algorithmic quality as well.
Any bounds are very, very high, though, way above human efficiency,
and likely enough to allow hard takeoff with any reasonable
This archive was generated by hypermail 2.1.5 : Wed Jul 17 2013 - 04:01:02 MDT