In practice, you can get away with this rule and do learning with less than 10 times the number of features in your data; this mostly happens if your model is simple and you are using something called regularization (addressed in the next chapter).

Jake Vanderplas wrote an article (https://jakevdp.github.io/blog/2015/07/06/model-complexity-myth/) to show that one can learn even if the data has more parameters than examples. To demonstrate this, he used regularization.