Re: AI: Relative difficulty

Eliezer Yudkowsky (
Tue, 21 Jan 1997 12:57:12 -0600

[Anders Sandberg:]
> How would you build a world where 1+1=3? It would be rather inconsistent,
> and thus unlikely to work (although the experiment could work just as
> well with a child in a inconsistent VR, of course). A better example
> would be a world with hyperbolic geometry or objects that are not persistent.

Of course it would be inconsistent! One and one *don't* *make*
*three*! Or at least WE think so. This would actually be best by an
iterative process: You bring up a child in a crude, hacked-up 1+1=3
universe, where the main manifestation is that any time the child takes
an object and brings it toward another object, a third object appears,
perhaps a cross between the other two objects. Then this child designs
a more thorough, elegant 1+1=3 universe by changing those aspects of
reality that intuitively appear inconsistent. After ten generations,
you might have a very, very, strange but cognitively (if not logically)
consistent Universe where one and one make three.

The question is: Can [1+1=3] *ever* appear *completely* consistent,
cognitively, thanks to adaptation by the child... or will our built-in
processes of visualization interfere?

--       Eliezer S. Yudkowsky

Disclaimer:  Unless otherwise specified, I'm not telling you
everything I think I know.