Hello @Bio_J,
I think you have pointed out the problem of measuring model performance with the training set! However, that is not the intention. Andrew is just showing us how to carry out the train/test procedure, and he will talk about how we use these two values, for example, when he discusses diagnosing Bias and Variance in one of the next videos.
So, for now, please don’t consider that we are free to choose either of the two cost values as model performance measurement.
Cheers,
Raymond