RE: Weaknesses in FAI

From: Ben Goertzel (
Date: Sun Oct 24 2004 - 07:12:08 MDT

> If what I ask has been beaten to death already, then please
> someone direct me to the appropriate archive threads.

It's been beaten to death already, but I don't have the archive indexed by
topic and I don't feel like searching it for you ;)

> But Ben, what I asked was basically what part of FAI is least
> supported by evidence, and taken most on faith? (I didn't want to
> say it so baldly because it sounds insulting.)

None of Eliezer's FAI theories are supported by evidence. They are all
basically supported by his conceptual, logical argumentation.

As Wei Dai pointed out in a recent message, Eliezer has made a lot of strong
conceptual arguments as to why creating Friendly AI is a hard problem. He
has outlined the problem very nicely. However, his proposed solutions to
this problem have not been convincing to very many people -- and he has
generally tended to reject his own solutions not too long after formulating
them. His current sketch of a solution -- "collective volition theory" --
seems both wildly impractical and severely frightening to me and many

> Should I
> understand you to believe that Eliezar may have some nice
> theories but has yet to produce code and validate himself?

It is certainly true that Eliezer has not published or described any
interesting software systems of his creation or conception. The only
software design he's made public (so far as I can recollect) was for the
Flare programming language, which I personally didn't care for much, but
which was anyway only peripherally related to FAI.

However, his explicit goal is to create a solid theory of FAI before coding
an AI. Furthermore, I believe he is now working with some software
engineers on implementing some of his ideas.

I think it is important to have smart people like Eliezer thinking hard
about FAI and related issues. And I have generally found Eliezer's writings
more insightful than those of other folks who write about these topics ...
even though he hasn't, in my view, come up with any real "answers,"
conceptually let alone code-wise.

-- Ben G

This archive was generated by hypermail 2.1.5 : Wed Jul 17 2013 - 04:00:49 MDT