I still can‘t get over learning just being parameter optimization. Like all possible forms of knowledge are already representable if only you knew the right weights. Technically that‘s probably true even for humans, but conceptually, it seems lacking to me.