Isabelle Levent

A recurring theme in participant feedback was that the language model lacked taste and intentionality...In contrast, good writers are skilled not only in producing but also discerning good language. In other words, they have taste, the ability to decide why one sentence is interesting while another is not.

Crawford elaborated in an interview. “When you have this enchanted determinism, you say, we can’t possibly understand this. And we can’t possibly regulate it when it’s clearly so unknown and such a black box,” she says. “And that’s a trap.”

Allen feels the same way. “I believe if you focus on the negative with AI, then that will come true,” he says. “And if we get more people focusing on the good and positivity that we can do with it, then that will come true.”

What’s difficult is to state our aesthetic values clearly enough to enable the program itself to make the evaluation at each generation.

We find that models learn just as fast with many prompts that are intentionally irrelevant or even pathologically misleading as they do with instructively “good” prompts. Further, such patterns hold even for models as large as 175 billion parameters (Brown et al., 2020) as well as the recently proposed instruction-tuned models which are trained on

Many methods for creating these models don't (and to be honest can't) attach the name, website and other details of every image and piece of text used to a create a new image in the metadata to every step of the process.

There’s another edge case as well; in theory, with the same prompts and the random seed that’s used for generating the images, you could end up with someone else generating the same, or a very similar, image as what you created.