Re:New AI edits images based on text instructions (instructP...
by TeMPOraL on Sunday January 22, 2023 @07:00AM
(#34476524)
I don't get it. Pre-ChatGPT prompt engineering was a BS exercise in guessing how a given model's front-end tokekizes and processes the prompt. ChatGPT made it only more BS. But I've seen a paper the other day, implementing more structured, formal prompt language, with working logic operators implemented one layer below - instead of adding more cleverly structured English, they were stepping the language model with variations of the prompt (as determined by the operators), and did math on probability distributions of next tokens the model returned. That, to me, sounds like valid, non-BS approach, and strictly better than doubling down on natural language.
[ Parent
| on HN ]
Re:New AI edits images based on text instructions (instructP...
by c7b on Sunday January 22, 2023 @07:36AM
(#34476772)
Think about the problem in an end-to-end fashion: the user has an idea of what sort of image they want, they just need an interface to tell the machine. A combination of natural language plus optional image/video input is probably the most intuitive interface we can provide (at least until we've made far more progress on reading brain signals more directly).
How exactly we get there, by adding layers like on top like language models, or adding layers below like what you described, doesn't seem like such a fundamental difference. It's engineering, you try different approaches, vary your parameters and see what works best. And from the onset, natural language does seem like a good candidate for encoding nuances like "make it pink, but not cheesy" or "has the vibes of a 50's Soviet propaganda poster, but with friendlier colors".