RE: How hard a Singularity?

From: Ben Goertzel (
Date: Sat Jun 22 2002 - 15:40:08 MDT

RE: How hard a Singularity?

There are no really well-fleshed-out interspecies intelligence measures!

Of course, one could try to make some up, and I've done some work in this
regard ... but the details of such an endeavor quickly become highly

In this case, I think that Eliezer and I actually have a common
understanding of what "superhuman intelligence" means, even though we don't
have an agreed-upon rigorous definition.

One interesting intelligence measure would be the speed with which a system
can prove a certain set of theorems given a certain set of axioms. On the
other hand, mathematical intelligence is only one of many kinds. A more
serious intelligence measure would involve measuring the rate at which a
system can solve a very diverse collection of tasks with known answers
(including some relatively easy ones and some very hard ones ). The choice
of tasks is somewhat arbitrary, reflecting the fact that "intelligence"
itself is not a precise mathematical concept, but rather an intuitive,
natural language concept, with a nebulous and shifting meaning.

-- ben


> I think the period of transition from human-level AI to superhuman-level
> will be a matter of months to years, not decades.

  I suppose I could see a month, but anything longer than that is pretty
  to imagine unless the human-level AI is operating at a subjective slowdown
  of hundreds to one relative to human thought.

  ### I am exasperated by the lack of rigour in both of the above
statements - neither one refers to a clearly defined measure of
"superhumanity" in intelligence. I have a challenge for all - try to come up
with a way of quantifying what you mean here. The statement that it'll take
a month from HL-AI to SAI is trivially true - if the HL-AI is worth its
electricity, it *will* become superhuman in a month, but will it be
superhuman enough to make a difference?

  No doubt there are many ways of quantifying intelligence, in terms of
absolute performance on standard tasks, in relation to other intelligences,
by the degree of comprehensibility of its thought/concept/percept content -
using almost any of them will make the discussion more interesting.


This archive was generated by hypermail 2.1.5 : Wed Jul 17 2013 - 04:00:39 MDT