Tech-utopia was often portrayed (by enthusiasts) in the past as one where the robots took care of the mundane boring tasks, leaving the human masters to sit in leisure and pursue their creative artistic endeavours.
From that background, it's ironic that some of the first successful mass applications of AI have been in fiction-writing and art, which are among the first things people think of as creative endeavours.
@digital_carver
Fiction writing?
I've seen some script enacted and read a few things made by AI (deep neiral networks, more precisely), and they're, shall we say ... abstract?
There's a lot in image recognition/synthesis, although the results are often still not quite there. OTOH, Chess and Go are firmly in "AI" hands, of course. Otherwise, NNs seem to be mostly in the "quick decision" category: modelling stuff that doesn't have to be 100% but is too difficult or slow without NNs.
@Mr_Teatime
True 😀 I was thinking in terms of availability (i.e. everyone could try them out), but in terms of results they're not quite on par with the Chess or Go AIs. People have posted impressive specimens from AI Dungeon and so forth, but those are the rare fully-coherent ones among the many "abstract" ones as you put it 😁
What kind of NN applications do you have in mind that's in the "quick decision" category?
@digital_carver
I came to NNs from the angle of respone surfaces (technically, curve fitting counts as "machine learning"...), and they basically allow a trade-off between accuracy, evaluation speed and inpit data volume that was not possoble before.
I mostly did Kriging and similar "Bayesian" modelling of noisy data, but once you have a certain amount of (nonlinear) data, NNs become hard to ignore. Effectively, image synthesis is "interpolating" images, too.
@digital_carver
hmm... now that I think of it: How come there is no NN-based video compression yet? You train a NN on the frames of the video, and then it "predicts" the look of the video at any point in time... with the right choice of neuron types, you could even get slow-motion interpolation for free.
@digital_carver
A colleague was using them for response surfaces (produce training data with complicated expensive model, then use NN as predictor in optimisation loops), and there's of course real-time image recognition, where programming the thing directly is hard and would take ages (if possible at all). Same thing with chess or go: explicit programming takes ages, but NNs can usually pick a move good enough to beat a human, even if you're never sure it's the perfect move.