While I get the underlying sentiment, it's worth noting that there have been examples of success, e.g. Deepmind with AlphaFold for protein folding with relatively little domain knowledge.
AlphaFold not only depended on a great deal of domain knowledge, it actually used one of the most famous physics-based methods. The only people who think that AlphaGo was made by a bunch of noob AI guys with a neural network are the people who don’t know anything about it.
I said relatively, they certainly had less domain knowledge than the competing teams. I have always been a strong proponent of combining ML and NNs with traditional methods.