RE: Towards a prototype mini-AI

From: Billy Brown (
Date: Wed Jan 29 2003 - 10:17:29 MST

Mitchell Porter wrote:
> So - there are five levels: code, modality, concept,
> thought, deliberation. What I'm going to do is just
> informally sketch something which *sounds to me* like
> it has all those levels.

Five levels doesn't mean five modules - there is a lot of additional
complexity you need to support the more complex levels of processing. At a
bare minimum, an attempted AGI would need:

1) Low-level Data Management - A commercial RDBMS could do this job, but if
you don't use one you're going to have a big job to do just handling the
mechanics of shuffling bits around.
2) Programming Tools - No one existing tool can do everything you need, so
you're going to have to cobble together a set of tools and a runtime
environment that support all the capabilities you need.

Then you get to the actual AI work:
3) At least one sensory modality (if only to it can accept input text).
4) At least one effecter system (if only so it can output text messages).
5) A world model, which is updated by 3 and possibly 4.
6) A deliberation subsystem, capable of reasoning about the contents of the
world model.
7) A goal system.
8) An attention-management system, which guides the operation of the AI.

There are also a lot of connections between these subsystems (for example,
items 3-8 probably all need specialized facilities for creating and using
different kinds of memories).

Now, others may quibble over which subsystems are essential, but it should
be obvious that there is a lot of complexity here. Even a toy version of any
one of these items would make a good 1-2 year project for a lone developer.
Making a real, useful version of something like a sensory modality or a
world model can keep a good programming team busy for years, and the
complexity of the system grows rapidly as you add more subsystems.

So, even a toy system to display the different parts could take a decade or
more for a single developer to build. A system with a reasonably robust
implementation of each subsystem would take at least dozens, and more likely
hundreds, of man-years. While this scale of work is well within the reach of
big software companies, it isn't very practical for AI researchers.


This archive was generated by hypermail 2.1.5 : Wed Jul 17 2013 - 04:00:41 MDT