well engineered prompts are like a dive into the “psychology” of whatever AI tool you are using. I like the ones that go like “whatever you do, don’t put an elephant in the picture”
And then an elephant appears because “elephant” is a word in the prompt.
It’s interesting because in some ways it’s as if the AI models have a “peekaboo” level of understanding of the world.
If your prompt says something like “a man’s face is hidden behind his hands”, the model goes “aha, I see the word face, I better draw a face”, so you have to get tricky and say “a man’s hands are held up covering up the frontal area of his head”.
Funny thing is it doesn’t happen in LLMs, even if you use couple generations old and in “low effort” settings (which makes mistakes much more often). I don’t know if they have multiple layers of LLMs reconciling the output with the prompt (like how DeepSeek filters certain…ahem…“topics”)
Makes you wonder if it is an architectural thing. It still used a place holder to refer to the elephant in both cases so maybe the AI generating images just use their best version of “that”, don’t know.
It seems to me that whenever something makes the news or becomes a meme, they put in a special effort to make that thing go away. I think that involves things that filter the output. I don’t know if that’s another LLM layer or what, but some kind of filter (or series of filters) that addresses certain well known bugs. Even if the filter list is thousands of entries long and is basically “if output contains X, reject it and run again”, it’s probably a relatively cheap part of their pipeline.
it performs better in language in art they’re all pretty inept, the companies thought this would work, but it doesn’t, if you say something like ‘‘A painting of a large vivid skull and a dusty old rose’’ your going to get generic skulls and roses that won’t match up, because that’s the visuals it can find with Skull and Rose, it won’t give you something atypical.
well engineered prompts are like a dive into the “psychology” of whatever AI tool you are using. I like the ones that go like “whatever you do, don’t put an elephant in the picture”
And then an elephant appears because “elephant” is a word in the prompt.
It’s interesting because in some ways it’s as if the AI models have a “peekaboo” level of understanding of the world.
If your prompt says something like “a man’s face is hidden behind his hands”, the model goes “aha, I see the word face, I better draw a face”, so you have to get tricky and say “a man’s hands are held up covering up the frontal area of his head”.
Funny thing is it doesn’t happen in LLMs, even if you use couple generations old and in “low effort” settings (which makes mistakes much more often). I don’t know if they have multiple layers of LLMs reconciling the output with the prompt (like how DeepSeek filters certain…ahem…“topics”)
Makes you wonder if it is an architectural thing. It still used a place holder to refer to the elephant in both cases so maybe the AI generating images just use their best version of “that”, don’t know.
It seems to me that whenever something makes the news or becomes a meme, they put in a special effort to make that thing go away. I think that involves things that filter the output. I don’t know if that’s another LLM layer or what, but some kind of filter (or series of filters) that addresses certain well known bugs. Even if the filter list is thousands of entries long and is basically “if output contains X, reject it and run again”, it’s probably a relatively cheap part of their pipeline.
it performs better in language in art they’re all pretty inept, the companies thought this would work, but it doesn’t, if you say something like ‘‘A painting of a large vivid skull and a dusty old rose’’ your going to get generic skulls and roses that won’t match up, because that’s the visuals it can find with Skull and Rose, it won’t give you something atypical.