I love the premise of this piece: AI’s ultimate value is not prediction, but provocation—helping us rethink the foundations of science. I like the way this article invites to reverse the value of how we do AI: if a model fails to predict a label, it's not that the model is bad, it may be that the label is based on reductionist categories. One thing that I find a little questionable: yes, we don't understand how our brain works, but we have a sense that we understand the world with the same abstractions that the article is looking for in AI. But we don't organize these abstractions from the inner working of the brain (spike trains or fMRI representations) as the article invites us to find the new abstractions in the clustering of the embeddings... I am not sure that the embeddings have any reality or "truth" to them. At any rate, this article is a wonderful and provocative invitation to rethink AI from a fresh perspective.
I love the premise of this piece: AI’s ultimate value is not prediction, but provocation—helping us rethink the foundations of science. I like the way this article invites to reverse the value of how we do AI: if a model fails to predict a label, it's not that the model is bad, it may be that the label is based on reductionist categories. One thing that I find a little questionable: yes, we don't understand how our brain works, but we have a sense that we understand the world with the same abstractions that the article is looking for in AI. But we don't organize these abstractions from the inner working of the brain (spike trains or fMRI representations) as the article invites us to find the new abstractions in the clustering of the embeddings... I am not sure that the embeddings have any reality or "truth" to them. At any rate, this article is a wonderful and provocative invitation to rethink AI from a fresh perspective.
Great article Sean. Unless we follow many of your suggestions AI is in danger of just producing a good deal of expensive distracting noise or echos.