Post #2790162

@Moon@shitposter.club @DrewKadel@social.coop @PCOWandre@jauntygoat.net @thatguyoverthere@shitposter.club conversation she talks about putting humans into the loop to help with the current stumbling block, defining where and how to abstract ("when you're looking for a tumour, pay attention to *this* in the scan and not *that*). similarly, when a model has the pieces and rules of chess predefined for it it then can out-perform human grandmasters, showing this really is the current limiting factor humans have got a lot of these abstractions predefined for us as well in our embryonic-developed architecture (e.g. facial recognition hardware and whatever in practice constitutes chompy's "language acquisition device"), but the algorithmic approach we use is clearly also much more aggressive at abstracting on the fly, and we'll need to adjust the training of models to do the same