News for hackers. Stuff that matters.

A comment on Things are about to get a lot worse for Generative AI.

Re:Things are about to get a lot worse for Generative AI by intrasight on Saturday December 30, 2023 @07:56AM (#38814940)
Just make LLMs be like your average human and forget details. I know that it's easier to say than to do, but so are many things worth doing. I can't plagiarize - my language and visual memory doesn't work that way. Such an LLM will have to "create" and answer from more fuzzy memory.
[ Parent | on HN ]
Re:Things are about to get a lot worse for Generative AI by qolop on Saturday December 30, 2023 @08:21AM (#38815078)
The class of models that Yann Lecun is bullish on (look up I-JEPA) do exactly this.
[ Parent | on HN ]
Re:Things are about to get a lot worse for Generative AI by intrasight on Saturday December 30, 2023 @11:13PM (#38821720)
But I-JEPA is non-generative. It does semantic image interpretation.

Okay, I guess it is related as my brain only does semantic image interpretation. (edit: my brain can create images, but only when I'm unconscious)

So with such a model, if you ask it to create an image, it would first create a semantic grammatical model of what you had asked for, and then perhaps draw it with colored pencils. I sort of like that. It's all that I could do. And it would be unlikely to violate any copyrights.

[ Parent | on HN ]