[ExI] Automated black-box-based system design of unsupervised hyperintelligent learning systems

[Adrian Tymes]

2011/9/18 Amara D. Angelica <amara at kurzweilai.net>:
> Does anyone know of a system for automated black-box-based system design of
> unsupervised intelligent hyperintelligent learning systems?

If it is unsupervised, how does anyone (including itself) know it is learning?

You can record all the data you want, but without some method of generating
and validating hypotheses, it's just data.  And that method is usually where
the "supervision" comes in.

The only existence proof of a completely unsupervised evolution has had a
runtime in the billions of years.  Most simplifications that allow emulation of
reality at faster than real time, eliminate one or more factors that this
process used to determine which organisms reproduced and how often.

In order to, say, take a random character generator and evolve it toward
something that can output English, you need an algorithm that can
determine whether "abfdslkjlr" or "fslweo.pds" is better, because that's
about the coherency of output you're going to get for the first several
generations.  Unless, that is, you're able to screen 38^140 or so
candidates per generation (letters, numbers, space, and period, for a
Twitter post) - which is analogous to how that existence proof did it.