paarth

One problem with the argument that “AI can’t generalize” is that humans also can’t generalize. We mostly just train on the test set. This isn’t just conjecture, mind you. It’s proven. There’s a famous philosophical problem, Molyneux’s problem. It asks if blind people, who identify 3D shapes like cubes or cylinders by touch, would be able to, if cured of their blindness, recognize those shapes by sight alone. The answer is no. In the early 2000s, we cured some people blind from birth and gave them that exact test: identifying 3D shapes by sight which they’d previously felt. The subjects were helpless to connect the cube they’d once held with the cube they were seeing, helpless to differentiate it from even a sphere. This explains why it’s so rare to see humans, even our best and brightest, making genuinely deep connections. We can’t. We don’t have the architecture, the power. We’re remarkably primitive.