Chris Olah's Universality Hypothesis states that different neural network vision models converge on identical feature representations regardless of architecture or training data