I built a custom GPT and uploaded photos to the knowledge base with the assumption it would share these, if asked.
First test result was that a DALL-E generated image would always be shared, so I disabled the DALL-E action and retested.
Result: ‘sorry, I can’t retrieve that image’. My guess was that the image size was too large, so I reduced to a 35kb jpg.
Retest results:
" Let me share a picture.
[image description]
Unfortunately, it looks like I’m unable to retrieve the image right now. I wish I could show you, but…"
That is, it describes an image from the knowledge base, rather than show/share it.
When I think about it, directly sharing anything in the knowledge base is dangerous. The custom GPT could share malicious code, NSFW pics, etc., so this outcome shouldn’t surprise me. However, I feel like anyone who tries to develop a custom GPT is flying completely blind without proper documentation, if it even exists.
I get the rush to market, ‘let them be the Guinea pigs’ aspect of things, but it’s difficult to support a product where the rules are almost entirely ambiguous, because we lose time, energy and resources just trying to figure the game out. I hope this changes.