Re: Floppy take-off

From: James Higgins (
Date: Wed Aug 01 2001 - 18:48:25 MDT

At 05:12 PM 8/1/2001 +0200, you wrote:
> > [...] write an AI tool that only knows how to improve code.
>"Improve"? I can see how you could use GA to optimize an algorithm,
>but to make it more intelligent? (You can't possibly mean an AI tool
>that opmitizes code, since that already exists and would imply that
>the Singularity should already have happened.) Are you saying that we

Not so, since there doesn't exist any software for general intelligence to
base improvements on! I think Gordon is suggesting something that would
make educated changes, test them, and then only keep changes that actually
improve the performance of the AI. This would be a very slow and tedious
approach, but it could potentially improve the AI to some degree.

>should use an artificial "dumbass" (our newly coined technical term),
>aka an idiot savant that is a wiz coder, to make more intelligent
>code? That would basically contradict Higgins, who said "others on
>this list have argued, quite convincingly, that it would at least have
>to be intelligent enough to understand what it was doing". I agree
>with him. How would the AI tool know whether it's stumbled upon
>something intelligent?

Well, if the idiot savant just happened to be a whiz at the exact tasks
required to improve the AI, then I suppose it would work. Think of the
above as advanced, automatic parameter tuning. It couldn't (I would
imagine) come up with radically new and improved AI algorithms, but it
could fine tune the ones being used. Which could provide just enough boost
for the AI itself to begin the takeoff.

>- Joaquim Gāndara

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