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

spike at rainier66.com spike at rainier66.com
Sun Apr 16 18:52:14 UTC 2023



…> On Behalf Of Jason Resch via extropy-chat
Subject: Re: [ExI] e: GPT-4 on its inability to solve the symbol grounding problem




>…It "grounds" (I put that in square quote because there is never any direct connection with the objects themselves) the meaning to the patterns inherent within the set of symbols themselves. …Jason



Excellent observation.  From that concept, we can swing directly to a plausible explanation for how ChatGPT can be trained to draw things.


Suppose ChatGPT has access to billions of capchas, solved correctly by humans, such as this one:



Given those, ChatGPT wouldn’t need to understand exactly what is a helicopter, but would only need to recognize a digital pattern, the bitmap image of a helicopter.  With that, we can imagine GPT figuring out common elements in those images and can produce a drawing of a helicopter, even if it is a bit wonky.  Perhaps it looks like what happens 


I don’t have DALL-E credits.  Does anyone wish to use their account to see if DALL-E can draw a helicopter?



-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.extropy.org/pipermail/extropy-chat/attachments/20230416/664bf818/attachment.htm>
-------------- next part --------------
A non-text attachment was scrubbed...
Name: image001.jpg
Type: image/jpeg
Size: 6284 bytes
Desc: not available
URL: <http://lists.extropy.org/pipermail/extropy-chat/attachments/20230416/664bf818/attachment.jpg>

More information about the extropy-chat mailing list