Re: DGI Paper

From: Eliezer S. Yudkowsky (
Date: Sun Apr 14 2002 - 21:54:36 MDT

Michael Roy Ames wrote:
> Ben wrote:
> > You omit to mention that thoughts as well as concepts can be remembered.
> > Much of episodic memory consists of memories of thoughts! Thus, thoughts
> > are not really "disposable one-time structures" as you say. Most but not
> > all are disposed of.
> Ben: I took Eliezer's DGI design intention, to define thoughts ==
> "disposable one-time structures", as a pragmatic deliniation between
> 'ephemeral' vs. 'permanent'.

Not just a pragmatic delineation between "ephemeral" and "permanent", but
also a systemic distinction between:

(a) the cognitive process that invents a thought,
(b) the active cognitive content that is the thought as it exists in working
(c) the cognitive subsystems that encode a thought as an episodic memory,
(d) the learned complexity that is the episodic memory of the thought.

So the thought itself may be a disposable one-time structure that fades out
of working memory, while the memory of the thought may be a permanent or
semipermanent episodic memory. The thought and its imagery fade; the
thought itself is nonrecurrent. The memory of the thought may live on as
recurrent learned complexity. But these are two different things with
different lifespans, different uses, and different representations and

As Roy Ames goes on to say:

> This 'cut' has to be made somewhere, as saving
> a detailed mind-state for every passing thought is just too expensive. I
> also have described this border as "a storm of thoughts raging over a sea of
> concepts"... when the storm of thoughts abate, the concepts remain, although
> somewhat mixed and changed.

(Nice metaphor!)

> This is not to say that a thought cannot be
> saved for later retrieval, just that it won't be saved in the same format as
> it normally exists. Instead it will be encoded as a memory - perhaps as a
> group of related concepts linked to specific imagery.

-- -- -- -- --
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