People believe that more parameters would lead to overfit instead generalization. The various regularization methods we use today to avoid overfit hadn't been discovered yet. Your statement is mostly likely about this.
Overfitting is the problem of having too little training data for your network size.