3 ms·

First, you pick the training data yourself. Consciously. It does not pick itself. More importantly, you pick the cost function. It's even harder to pretend it'

First, you pick the training data yourself. Consciously. It does not pick itself. More importantly, you pick the cost function. It's even harder to pretend it's picking itself. We consciously design it, whether you want to call it programming or not.

famouswaffles 3y ago
When the training data is essentially "all written text I can get my hands on", can't say you've made much in the way of conscious choice here. If the objective function is so vague that it allows you to complete any task then claiming conscious design also makes little sense. "I'm consciously designing you to do whatever you want!". Oh gee, that just makes everything so much better for some reason.
- vintermann 3y ago
  > When the training data is essentially "all written text I can get my hands on", can't say you've made much in the way of conscious choice here. Oh yes, that's a conscious choice. And not one which gets you a decent LM, incidentally. > If the objective function is so vague that it allows you to complete any task The loss function is not vague at all, and it certainly doesn't allow you to complete any task (it's more impressive that it allows you to compete any tasks at all, frankly). > I'm consciously designing you to do whatever you want! The point is that you aren't, and that "designing to do whatever it wants" is nonsense because by default it wants everything equally much / doesn't want anything at all (those are the same thing).