Petar discusses the computational limitations of neural networks, emphasizing the importance of aligning neural network components to learn linear functions for proper extrapolation. He highlights the need for implicit causal models to effectively extrapolate on various data sizes.
