current "#ai" = lossy search / compression incapable of semantic understanding
train on shit, receive statistical properties of shit. it don't care, it'll memorize anything
ask for bs and it'll give you some statistically average nonsense. matrix multiplication don't care.
this paper challenged current #llm models with questions they couldn't have seen before, and they scored 5%
with #stableDiffusion if you don't have a very average request or a trained supplementary model to support exactly what you're asking for, you get random nonsense instead
the state of the art of AI today just memorizes shit. we have not automated creativity yet. machines aren't thinking, yet
i really want this technology to be better, I want my #midnightBurger tangle. i'll even take the rude one. but we're about as far from that as we are from fusion power
