Re: Definition of strong recursive self-improvement

From: Mike Williams (
Date: Sun Jan 02 2005 - 13:59:54 MST

Are you suggesting that the AI is going to make random code changes and
then try to determine whether an improvement results? When I make a
change to a program, it's always with a purpose--improve speed, reduce
storage requirements, fix a bug, improve some particular functionality.
 After making the change, I evaluate the change against what it was that
I was trying to do (plus do other tests, such as regression testing to
ensure nothing else got screwed up in the process).

I agree that it's not trivial. But there are established procedures for
making changes to complicated software systems.
Mike W.

Russell Wallace wrote:
> On Sun, 02 Jan 2005 08:40:38 -0800, Mike Williams <> wrote:
>> Windows XP has system restore, databases have rollbacks, but a
>> super-duper AI can't recover from an undesirable code change? I'd think
>> the AI could do better than those systems: 1) backup self 2) install
>> mod 3) evaluate mod-self 4) decide not to keep the mod 5) remove
>> mod (or restore the affect parts) 6) _learn_ from the experience, and
>> not repeat that particular mod.
> "I think you should be more explicit in step 3" :)
> Seriously, this is precisely the problem. What algorithm do you
> propose the AI use for determining which of the two versions of itself
> is better?
> - Russell

This archive was generated by hypermail 2.1.5 : Wed Jul 17 2013 - 04:00:50 MDT