> Think of the problems AI has with creating appetizing food imagery, for example.
replyFor posterity, the memetic ugly food can be controlled and mostly filtered out with prompt engineering. Here's a quick ad hoc test example which (hopefully) uses the new model: https://chatgpt.com/share/6aebf030-9d38-83e8-a103-3509883905...
First image is fine, but attempts to make it more wacky turn it to the AI food slop. So then followup prompts deslopify it, mostly successfully.
Eh, looks like slop still in my opinion. The evenness and density of the teeth especially stands out to me as uncanny.
replytbh it got worse every time, the first one is the best lol
replydeleted
reply>This is not true. Think of the problems AI has with creating appetizing food imagery, for example. Images that every human annotator marks as unappealing with a high degree of consensus.
replyI think the way they'll solve it in the end is to create an ultra-high-fidelity simulation of the world including modelling taste and other senses on the level of physics, and then incorporate all of this into the training rewards.
I think this is how the machines know what tasty wheat tasted like.