Re: [sl4] How much do languages matter for AGI programming?

From: J. Andrew Rogers (
Date: Fri Nov 28 2008 - 13:18:54 MST

On Nov 28, 2008, at 11:24 AM, Martin Sawitzki wrote:
> As long as its turing complete....
> I guess Lisp/arc are good candidates, but also Java might be ok,
> depending on how complex the actual design gets.

It really depends on what, specifically, you want to make the language
do. Using LISP or Java or whatever is making a lot of assumptions
about what the problem will actually look like in implementation.
LISP has always been considered a classical AI language, but the
language is relatively poor for solving certain types of computational
problems and so by selecting LISP on the basis that it is ideal in
some fashion, one is make an assertion about the underlying nature of
AGI implementation itself. Those assumptions seem to never be backed
up by strong evidence. One could make an argument that the lack of a
very well-defined computational model for a complete AGI prescribes a
more generalist language like C++ that is capable of doing everything
quite well (while making nothing particularly easy).

Most AGI people seem to pick languages they like, and then rationalize
the choice with respect to its AGI optimality. In my opinion, it is
sufficient justification that one selects a programming language for
purely pragmatic and aesthetic reasons without any regard for its
nominal AGI optimality.

Most AI programming language arguments are little more than jousting
matches between incompatible self-deceptions.

> Developing the right language is just another subproblem of the
> whole AI business. When the theory is done you have a far better
> understanding what your language should be able to do.

There are many possible language and computational models, only a tiny
number of which ever show up as first-class models in a major
programming language. Therefore it is likely that AGI will not
directly map to any extant popular programming language, and the
syntactic sugar of an AGI optimal language will only be developed
after the fact -- when it will be largely irrelevant.


J. Andrew Rogers

This archive was generated by hypermail 2.1.5 : Wed Jul 17 2013 - 04:01:03 MDT