You need to remember that these LLM things aren't smart, they don't think, they aren't capable of reflection or semantic understanding. All it's doing is generating strings of text based on weighted probabilities. That's fine as long as you keep things very superficial, but if you try to get any depth out of it the gaps are going to become glaringly obvious very quickly.
You know how the "art" models can't draw hands? If you try to use these things for more than they're capable of you're going to get that kind of unnerving uncanny valley fuck up in your narrative. They can't draw hands because they're not capable of abstraction. They don't have an abstract understanding that a 'hand' is an object with certain characteristics no matter what position it's in. It can't do hands because they appear in so many different shapes in so many different contexts that it can't identify a consistent enough pattern to reproduce the hand. Faces are relatively consistent and easy, but anything that moves around too much - hair, hands, really almost any detailed object that's not a face, it can only draw very superficial approximations of. Jewelry is a good tell because there are so often pieces that don't connect to anything, or pieces that flow in to flesh or clothing.
It can’t do hands because they appear in so many different shapes in so many different contexts that it can’t identify a consistent enough pattern to reproduce the hand.
There are a lot of simple end runs around this glitch, even before the upgrades and refinements. Just requesting hands appear under an apron or behind the back, for instance.
Professional artists do shit like this all the time. There's a whole rant online about how Rob Liefeld can't draw feet. He's been churning out comics the old fashioned way (by stenciling over images he's not great at drawing) for decades.
All that is to say, we're not asking AI to work miracles here. Just to improve over what is rudimentary and generic digital conversations in a way programs like Replika have already proven successful at.
Yes, if you start too hard at the edges, you'll lose the illusion. But we're not trying to outwit Harrison Ford in Blade Runner, here. Humans are very easy to fool when they want to be fooled.
You need to remember that these LLM things aren't smart, they don't think, they aren't capable of reflection or semantic understanding. All it's doing is generating strings of text based on weighted probabilities. That's fine as long as you keep things very superficial, but if you try to get any depth out of it the gaps are going to become glaringly obvious very quickly.
You know how the "art" models can't draw hands? If you try to use these things for more than they're capable of you're going to get that kind of unnerving uncanny valley fuck up in your narrative. They can't draw hands because they're not capable of abstraction. They don't have an abstract understanding that a 'hand' is an object with certain characteristics no matter what position it's in. It can't do hands because they appear in so many different shapes in so many different contexts that it can't identify a consistent enough pattern to reproduce the hand. Faces are relatively consistent and easy, but anything that moves around too much - hair, hands, really almost any detailed object that's not a face, it can only draw very superficial approximations of. Jewelry is a good tell because there are so often pieces that don't connect to anything, or pieces that flow in to flesh or clothing.
The latest version of Midjourney is now able to draw hands with a fat higher consistency and quality. It only took them a year to polish that bit.
Case in point
There are a lot of simple end runs around this glitch, even before the upgrades and refinements. Just requesting hands appear under an apron or behind the back, for instance.
Professional artists do shit like this all the time. There's a whole rant online about how Rob Liefeld can't draw feet. He's been churning out comics the old fashioned way (by stenciling over images he's not great at drawing) for decades.
All that is to say, we're not asking AI to work miracles here. Just to improve over what is rudimentary and generic digital conversations in a way programs like Replika have already proven successful at.
Yes, if you start too hard at the edges, you'll lose the illusion. But we're not trying to outwit Harrison Ford in Blade Runner, here. Humans are very easy to fool when they want to be fooled.