In general, it’s good to restrict NN inputs to small range for the sake of faster converge. See this post for an example.
Let’s consider the 1st exercise in course 1 (housing price prediction based on number of rooms). Go ahead and set the ys to the actual price i.e. 50000 + num_rooms * 50000 and build the model, with a lambda layer say tf.keras.layers.Lambda(lambda x: x * 50000). Model loss is going to be nan during training.
I recommend standardizing features instead of using lambda layers or changing output activation of lstm to relu instead of tanh.