Anthony on Nostr: It's an important development for many reasons, but one is that there are many people ...

It's an important development for many reasons, but one is that there are many people now arguing something like this: because of double descent, we don't need to worry about the bias-variance tradeoff anymore. With enough compute we can continuously lower the bias, and with enough data we can continuously lower the variance. It takes a bit more subtlety to tease out why this is a reasoning error. It unfortunately meshes well with prevailing prejudices in thought, especially among technologists (the theory-free ideal, Han's notion of dataism, etc.), so it's a tricky argument to land. I like Brighton's notion of "the bias bias" for this: preferring ever-increasing model complexity really is a bias, unjustified by either evidence or theory.