Review of A2I2

From: Eliezer S. Yudkowsky (
Date: Sat May 04 2002 - 11:00:38 MDT

In reviewing Peter Voss's A2I2 I'm working with much less information than I
have for Ben Goertzel's Novamente. My review of Novamente is based on
reading the Novamente manuscript and having extensive conversations with
Ben; my review of A2I2 is based on seeing part of a demo at the Foresight
Gathering before I had to leave for the airport (aaargh!), which hopefully
was enough to give me a glimpse of the underlying architecture, plus various
bits of surrounding information, a few small exchanges with Voss, and so on.

Unlike Ben with Novamente, Peter Voss has not, as far as I know, expressed a
belief that A2I2's current architecture is adequate to general intelligence
or seed AI; Voss's philosophy of AI is that you start by building a mouse.
Okay. Not my personal approach but not an invalid approach either. After
seeing the demo my reaction was again that A2I2 was simpler than I'd
expected, but that could just be because it was a demo.

What I saw in the demo was an AI with a visual grid operating at multiple
scales of resolution, and a network of nodes in which each node represented
a visual pattern. When A2I2 runs across a visual pattern it hasn't seen
before - a visual pattern below a defined threshold of similarity to any
existing nodes - it creates a new node. Parts of A2I2 explicitly indicated
as being absent from the demo include A2I2's ability to issue motor actions
and a reinforcement system that links perceptions to actions. ("Visual
pattern" and "motor action" means A2I2 is bouncing around in the visual
environment of a 2D microworld, not that it's a robot.)

I guess my overall reaction is that this is fine as far as it goes, but I'm
not sure what Voss plans to do with it when it's finished. Build a better

-- -- -- -- --
Eliezer S. Yudkowsky
Research Fellow, Singularity Institute for Artificial Intelligence

This archive was generated by hypermail 2.1.5 : Wed Jul 17 2013 - 04:00:38 MDT