Moreover, the current Dalle UI already does that.

When you run a phrase, you get four images. Those images will stay in your history, but the ones you like you will save with the "save" button, so that they're in your private collection.

With this, you already have a great feedback system: saved - good, not saved - bad.

I've saved some of the worst images Dalle generated to be able to showcase just how bad it can be sometimes. And then other times the bad image is hilariously bad. They can probably build another layer on top of the feedback system though to filter that sort of thing out.

I would guess your use-case is a statistical anomaly. If most of the images that are saved are saved by people who like them best, which is most likely the case, enough data will erase the problem.

Doesn't the sample size for this have to be very large for it to make a difference? Genuine question.

With semi-supervised learning a small amount of labeled data, can produce considerable improvement in accuracy:

https://en.wikipedia.org/wiki/Semi-supervised_learning

https://towardsdatascience.com/semi-supervised-learning-how-...

Thank you!

Sure, but there are millions of people on the DALLE waitlist, who would happily rate the output for better performance / more credits. The famous ImageNet data set only has 1.2M images.