Re: Seed AI milestones - Explicit Knowlege Codification

From: Eugene Leitl (
Date: Thu Feb 28 2002 - 01:02:58 MST

On Wed, 27 Feb 2002, Samantha Atkins wrote:

> Fortunately, not all aspects of the human baby are essential to
> the learning machine. The trick is finding and recreating the
> minimum set in machina.

Unfortunately a lot of it is imitation learning, which would seem to imply
an anthropomorphic body, a (virtual) arena, and ersatz parents. Implies
lots of overhead, but the advantage is that you have a very human AI,
which has an opaque hardware layer (i.e. does not understand its own
coding, while having the classical advantages of halting, full inspection
of state, snapshots, etc).

The same applies to human uploads. A world full of uploads is much less
vulnerable to a runaway, both because they're not coupled to an ecosystem
(orbiting small boxes in space), and because their time frame of operation
is in the order of magnitude of an SI on computronium substrate.

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