> On the other hand, if DALL-E were to generate 10 billion images, and choose the best 2-3 itself and give those as output, and if at least one of those 2-3 wou

> On the other hand, if DALL-E were to generate 10 billion images, and choose the best 2-3 itself and give those as output, and if at least one of those 2-3 would be consistently great, then DALL-E could be indeed considered to be creating (good) art. It's worth noting that the OpenAI samples for DALL-E 1 used CLIP to rank generated samples, and got a big boost from that. For many model architectures, you can run them in reverse to do 'image -> caption', and 'score the caption' quality: if 'the caption is bad', that indicates your image was screwed up and low-quality (introduced by Cogview). DALL-E 2 doesn't use either approach, or finetuning on user choices like InstructGPT, and I dunno if OA is going to implement any of these, but there is a wide universe of techniques applicable here to improve quality and we should keep that in mind (https://www.gwern.net/Forking-Paths https://www.gwern.net/Forking-Paths) if we are going to make any assertions more sweeping than "this specific model, at this very instant, with this particular interface, is only at this level of quality".