> They don't give any scientific or mathematical answers.

Dead wrong. Predicting protein folding has been a scientific holy grail for decades, and neural network-based AlphaFold is the leader. No non-neural approach comes within a country mile of its performance. They created a fold database of unprecedented coverage which is being used in research.

He was decidedly not right. His claim was that neural networks are fundamentally not able to perform an XOR step, limiting any relevant computational capabilities. That claim was easily remedied later by slightly changing network architectures later on.