This question has come up before. I think you’re just over-interpreting the question. It’s never the case that just the architecture of your model is enough to guarantee success. You always need enough training data, you need the compute resources to run the training, you need to make sure the power grid stays up long enough to complete your training and on and on … But it’s not asking about any of those other requirements: it’s specifically asking about whether hand engineered features are required or not versus a complex enough model.