Will drawing AIs like DALL-E be able to draw convicing hands 9 times out of 10 before the end of 2024?
41
แน€5.1k
Dec 31
93%
chance

On January 1st 2025, I will prompt a drawing AI with the intention of making it draw pictures containing hands in various contexts, angles, lighting, etc.

I will grade each one as convincing based on my personal impression. I have a very pronounced natural disgust reaction towards deformed human shapes, to the point I can't stare at some drawings made by current AIs for more than 5 seconds without averting my eyes. So if I find I get that reaction from a hand I will grade it as not convincing.

To make the criteria a little more clear, nails coming out of the back of the hand, fingers coming out of the palm, slightly wrong proportions like fingers being too long, or the hand being too elongated wrt forearm, wrong number of fingers unless the hand is clearly cartoonish or an alien's, will definitely cause the disgust reaction.

Question will resolve to YES if 90% or more of the pictures, out of a sample size of 30, classify as "convincing".

EDIT 1: I commit to withdrawing any bet I made in this market by 15th December 2024.

Get แน€600 play money
Sort by:

how will you choose the AI though?

@1a3orn in order of priority:

1. The most common AI used to generate pictures of people FOR PROFIT (regardless of whether it's optimized on a specific type of content, as long as drawing convincing, realistic hands is relevant eg. adult content). The absolute prerequisite for this criteria is that someone is consistently profiting from content generated with this AI - profit is the best objective metric for usefulness I can think of
2. if there is an AI that is optimized on drawing realistic images of people, and the size of the FLOPS used to train it are at least 75% of the record of FLOPS used to ever train an image generation AI, I will pick that one
3. otherwise I will pick the image generation AI for which it took the highest amount of FLOPS to train so far

do you have any suggestions on how to improve this criteria?

predicts NO

another possibility would be to open a market when it's time to resolve the question, and ask something like "which AI is most likely to pass this test" or "which AI is the least likely to generate body horror when prompted as such"