A friend had the brilliant idea of asking one of those image AIs to draw a diagram of the anatomy of a human hand, since that seems to be the thing they have the most trouble drawing. The AI managed then produced these diagrams, which are frankly impressive in some ways, while 100 % confirming that they cannot draw human hands. Never mind getting the number of fingers right, look at these skeletal structures and tell me there aren't deeper issues at work.
Yeah, the imitation of the style of a medical diagram is actually very impressive. And as a human it is fascinating how it can be so good at imitating that particular style of drawing - which I think is very difficult - while not being able to draw the correct number of fingers - which I think is very easy - even on a single human hand drawn from straight above in the most neutral pose possible.
Yeah it's a fascinating demonstration of the limits of 'AI' in the fact that even with the full depth and breadth of humanity's knowledge, without clear and concise commands to direct it its about as useful as a paperweight.
I think they try to train these AI by having the people who requested the picture grade the results. That might help them learn to avoid making ugly mistakes like lines not meeting, but I think it will reinforce the problems in this picture. Because the only feedback this AI would get would be "FUCK YES! This is exactly what I wanted! Perfect drawing of a human hand!".
Yeah, like unless they program the bot parameters to interpret the vague language we use in common parlance into actual useful data that positively builds it's database then yet again it'll have the usefulness of a paperweight but with an interesting gimmick.
This is a cool result, and I enjoyed watching it, so it's not exactly useless. They just cannot improve accuracy from most people's feedback. I don't think I'm alone in giving the top right picture with seven fingers and a creepy spider-human hybrid skeleton top marks. And the AI will just absorb positive feedback as having done the right thing. This will not make the AI better at drawing human hands, but it might make it even better at drawing abominations. I'm there for that.
Same, actually. Its like having an AI Zdzislaw Beksinski.
It's because the difficulty of highly detailed illustrations lies in the mechanical aspects that stand in between the idea of the thing in one's brain and the abstraction of that on the paper/tablet/whatever medium. For these AIs, there's no idea of the thing or messy systems of tendons and muscles getting in the way, there's just a feverish image formed through data association.
The real breakthrough I think would by layering the system with something that could create that sort of idea of a thing, even if it's just a collection of barebones 3d models and a neural network trained to animate them that then provides reference images into an AI that turns that into art, sort of like a very fancy automated rotoscoping system. That would probably be enough to make them get things like number and position of fingers and limbs right, at least.