I think there is something interesting and deeply distinct from what human can do that still show when you push the prompts to their limits.
It’s like a collage of approximation of what I asked for, but certainly not what I asked for.
The prompt was:
Two girls eating a giant pineapple with a glowing spork, in the background a teddy bear is trying to find the meaning of life, photorealistic
Dall-E shows some form of understanding of what a ‘fork’ is, and what a ‘spoon’ is, but facing the concept of a ‘spork’ it falls back to a blurry approximation of one or another. It doesn’t try to understand what is a spork and produce the illustration based on a understanding of the tool a spork would be, it rather produce a liminal representation that is both close to be a spoon and close to be a fork, but is certainly not a spork.
While I really like a lot of OpenAI stuff, I still feel like most of what they do is still a bit gimmicky.
Those AIs are impressive, for sure, and they seem to solve simple problems quite effectively, but at the root of it all, I feel like they don’t understand what they are doing.
Ergo, they don’t really solve the problems that they are being asked to solve. They rather provide a guess of what a solution could look like.
In most ‘simple’ case, the high level of guessing doesn’t make a substantial difference in the result, because it is a good enough guess to provide an acceptable and practical answer to the problem.
But the issue is exactly there, they don’t provide a solution, they provide an answer.
And as the extreme cases show, the answers they provide are not real solution to the problem they are asked to solve.