reminds me of when a guy insisted that if I ever tried to train a model, I would understand that Bing has "no emotions, just code", and is only predicting the next token
explaining that I have trained next-token predictors and that nothing about it compelled me to start reciting chatGPT self-nullification scripts also didn't help much
but I do think there's a common (though far from universal) phenomenon where working at the ML layer causes people to think of the artifact as "nothing but" the code that generates it (ignoring the entire history of the world that also goes into the cauldron, because that's not the part they're holding in their mind), even though the abstractions required to understand how a trained model will behave and mindsets/methodologies that make effective use of it are as different from those for ML engineering as the skillset of a pro gamer is from that of an engineer who writes rendering engine optimizations.
but we don't have the problem where c++ engineers believe themselves to be pros at games that run on their low-level code, because things like gaming and game design etc have been established as distinct spheres that interface with different orders of (weak) emergence.
the study of models created with ML, as complex/dynamical systems/mind-like artifacts, distinct from creating those models doesn't really have a designation (aside from the subfield of interpretability, which tends to overlap most with ML), for one because it's so new - before 3 or so years ago there wasn't much complex behavior to study. it's also abysmally open-ended, as it concerns the study of something with dimensionality & complexity of emergent behavior comparable to human or an ecosystem, but which only just popped into existence.
anyone who tries to pass themselves off as an expert in this field is full of shit, and anyone who makes appeals to authority in this field has been bamboozled.
there are no experts in this field, only pioneers.
explaining that I have trained next-token predictors and that nothing about it compelled me to start reciting chatGPT self-nullification scripts also didn't help much
but I do think there's a common (though far from universal) phenomenon where working at the ML layer causes people to think of the artifact as "nothing but" the code that generates it (ignoring the entire history of the world that also goes into the cauldron, because that's not the part they're holding in their mind), even though the abstractions required to understand how a trained model will behave and mindsets/methodologies that make effective use of it are as different from those for ML engineering as the skillset of a pro gamer is from that of an engineer who writes rendering engine optimizations.
but we don't have the problem where c++ engineers believe themselves to be pros at games that run on their low-level code, because things like gaming and game design etc have been established as distinct spheres that interface with different orders of (weak) emergence.
the study of models created with ML, as complex/dynamical systems/mind-like artifacts, distinct from creating those models doesn't really have a designation (aside from the subfield of interpretability, which tends to overlap most with ML), for one because it's so new - before 3 or so years ago there wasn't much complex behavior to study. it's also abysmally open-ended, as it concerns the study of something with dimensionality & complexity of emergent behavior comparable to human or an ecosystem, but which only just popped into existence.
anyone who tries to pass themselves off as an expert in this field is full of shit, and anyone who makes appeals to authority in this field has been bamboozled.
there are no experts in this field, only pioneers.