[ExI] e: GPT-4 on its inability to solve the symbol grounding problem

BillK pharos at gmail.com
Sun Apr 16 20:58:12 UTC 2023

On Sun, 16 Apr 2023 at 21:47, spike jones via extropy-chat
<extropy-chat at lists.extropy.org> wrote:
> >...Thx BillK, I meant can it draw a helicopter, analogous to what Giovanni was doing with the apple experiment.  Those are two completely different experiments.  Giovanni what software did you use to get those apple/plate drawings please?  spike
> Reasoning:  the entire grounding problem thread (very enlightening it is, thx to all for the free education) is all about the level of understanding software is displaying.  It is easy enough for me to imagine taking a jillion capchas asking for the BI to click all the squares with a helicopter, create a photographic image from that, label it "helicopter" and drop it into a huge lookup table.
> But... if it takes all his images labeled helicopter and draws a helicopter, that to me is demonstrating a whole nuther level of understanding.  It is a much more difficult task, one that most humans cannot do with any aplomb.  For instance, right now, draw a helicopter.  Don't look at photo, just draw one.
> Don’t draw an open-frame like this one, but rather a different concept please, so you aren’t influenced by my amateurish attempt.
> The reason I find this experiment interesting is that I do know how a helicopter works, know hte parts of the very impressive mechanical control system, the stabilizer bar, the swash plate, the engineering magic that keeps a helicopter flying choppy side up, I know all the reasons why if you knew how a helicopter control system works you would flatly refuse to get aboard one, even if it is a medivac and you have a severed artery.  You would tell the pilot no thanks, I will wait for the four-wheeled ambulance and eat lots of spinach until it gets here.
> I want to see what an AI will draw if asked because most of its digital images of a helicopter don’t show any of that and tell it anything about all that control stuff.
> spike
> _______________________________________________

The AIs don't work like that.  AIs like DALL-E try to produce
photo-realistic images.  Depending on the request, they may be very
surrealistic, but life-like and colourful.
They don't do line drawings.


More information about the extropy-chat mailing list