When AI hype has settled some, I'd like to see neural primitives be considered to be part of standard CS education along with other ADTs.
Hype makes ML look like too good to be true magical algorithms, and then fails because it was a grift all along. But, there is legit value.
When you watch educational videos on auto encoders, U nets, etc etc, they talk about specific things they are good at to fit in a larger solution.
@demofox I think it would also help if all the ML stuff didn't feel like it was walled off in its own incompatible ecosystem. It's all jupyter notebooks, models I don't know how to work with, and cloud APIs when I just want to do things like "make a C++ library (which runs on my computer) that generates plausible rgb tuples from the mouth feel of words". No idea how I'd even begin to do something like that.
@demofox it often feels like all the "AI" projects are made by people that either want to bleed me dry or want nothing to do with people like me, but I think some of that is just because the field has it's own deeply entrenched norms and established tools and workflows and jargon that is significant enough for an outsider to be bewildered when trying to navigate it all
It'll be a good sign when people craft algorithmic solutions to problems, where some of the components are traditional ADTs, and some are ML components, each used in appropriate ways towards a holistically well engineered solution.
ML can do some interesting and useful things, but not every problem is best solved as a black box, or works best when abstracted to matrix multiplies.
@demofox do you think the primary value of adding these models to the 'standard' education would be in direct awareness of them (anticipating moments like "aha, we could use an autoencoder here"), or a more generalized familiarity with the basic bag of tricks that make up the current world of convnets (i.e. come up with some suitable architectural design and throw gradient descent at it)?
@SnoopJ I was thinking more of the first. The second is useful too, but maybe differentiable programming should be it's own course. Making curriculum is hard!
Add comment