RE: Microsoft aflare

From: Ben Goertzel (
Date: Wed Feb 27 2002 - 06:51:41 MST


> For the record, I claim AI is a large scale massively numerical (not
> necessarily float, in fact a I would advice to stick to integers strictly)
> crunch application.

Although the inner loops are always numerical operations, even so, the
arrangment of numbers in appropriate ways, and the appropriated ordering
of numerical operations, requires lots of non-numerical data structures,
which are not simply arrays. So calling AI "number crunching" doesn't
tell very much of the story. You could just as well call it "tree
traversing" for example... equally incompletely...

>Plus, bootstrapping resources are rather excessive, as
> you have to necessarily traverse more barren fitness landscapes first.
> Given that your first implementation will be based on a commodity cluster
> (possibly, commodity cluster with a few kNodes), I don't see how you can
> evade using ANSI C and MPI as the implementation layer, and be completely
> agnostic about elementaries like internode bandwidth and messaging
> latency.

Although we are using C++ (gcc on linux) for our own work, it's easy for
me to see how another choice would be reasonable.

The choice of language for a project is a choice of a lot of different
and hard to meaningfully talk about in such general terms.

For direct source-code self-modification, there's a decent argument for LISP
as opposed to C++, as the program language semantics is more easily

Our choice of C++ is because we are valuing performance above all for this
of our system, because we've realized that thorough experimentation with
variants of
the system requires efficiency even MORE than a finished system will. gcc
kicks butt
on a Linux cluster of SMP machines...

But when we get to the stage of direct source-code modification (probably a
years at the present rate) we might re-code in LISP, or now, who knows.

Even though we made an error by choosing Java at Webmind inc., I still
that prog. lang. choice is among the least of the issues confronting would
be seed
AI developers. It's the mind design that's the obstacle, believe me.... I
think I have
a good one now but it's taken many years, and it's not simple enough for my
taste, although
we're proceeding with it until/unless something better comes along.

-- Ben G

This archive was generated by hypermail 2.1.5 : Wed Jul 17 2013 - 04:00:37 MDT