You would at least be able to say: No, we won't train an LSTM on 5000 datapoints (unless there's some heavy transfer learning going on).

Once you get your hands dirty, and need to debug unexpected issues, understanding what happens under the hood becomes more important.