RE: SIAI's flawed friendliness analysis

From: Ben Goertzel (
Date: Fri May 30 2003 - 14:47:53 MDT

> > This is where a pure reinforcement learning approach *may* be more
> > dangerous than a mixed-cognitive-methods approach like Novamente...
> I think consciousness and cognitive functions are part of a
> robust solution to reinforcement learning. They help minds
> create the simulation model of our very general world, in
> order to solve the credit assignment problem. Of course, when
> you are designing a system like Novamente, you may place them
> in different modules. But that doesn't change the fact that
> cognitive methods are necessary for effective reinforcement
> learning in a complex world.
> ----------------------------------------------------------
> Bill Hibbard, SSEC, 1225 W. Dayton St., Madison, WI 53706

I agree with your statements above.

But my point is that, given any fixed set of (behavior, reinforcement-value)
pairs, there are a LOT of different mental structures/dynamics that will
solve the problem of generating behaviors that maximize reinforcement-value
almost equally well.

Among these are, most likely, some that are going to be scary
Friendliness-wise and some that are going to be morally sound.

If this is the case, then the reinforcement-learning perspective is not in
itself adequate for discussing/analyzing Friendliness...

-- Ben G

This archive was generated by hypermail 2.1.5 : Wed Jul 17 2013 - 04:00:42 MDT