I wasn't expecting GPT-4 to be able to correctly answer "What is funny about this image?" for an image of a mobile phone charger designed to resemble a VGA cable - but it can.
(Note that they have a disclaimer: "Image inputs are still a research preview and not publicly available.")
I would bet good money that when we can test prompting with our own unique images, GPT4 will not give similar quality answers.
I do wonder how misleading their paper is.
They literally sent it 1) an a screenshot of the Discord session they were in and 2) an audience submitted image
It described the Discord image in incredible detail, including what was in that, what channels they subscribed to, how many users were there. And for the audience image, it correctly described it as an astronaut on an alien planet, with a spaceship on a distant hill.
And that image looked like it was AI created!
These aren't images it's been "trained on".
There's easily a 10:1 ratio of "it doesn't understand it's just fancy autocomplete" to the alternative, in spite of published peer reviewed research from Harvard and MIT researchers months ago demonstrating even a simplistic GPT model builds world representations from which it draws its responses and not simply frequency guessing.
Watch the livestream!?! But why would they do that because they already know it's not very impressive and not worth their time outside commenting on it online.
I imagine this is coming from some sort of monkey brain existential threat rationalization ("I'm a smart monkey and no non-monkey can do what I do"). Or possibly just an overreaction to very early claims of "it's alive!!!" in an age when it was still just a glorified Markov chain. But whatever the reason, it's getting old very fast.
What is funny is neither GPT-4 nor the host noticed that (or maybe the host noticed it but didn't want to bring it up due to it being "inappropriate" humor).
anything not on your list
