The article says that dynamically typed code and statically typed code has a similar number of bugs per line.

It also says that a software feature only needs 1/2.5 lines if you are using dynamic typing.

That means dynamic typing reduces bugs by 2.5x compared to static typing per software feature.

None of these are anywhere in the linked article: "2.5" "similar"

There is more than one study on the topic. But you can just take a look at the graphs.

Take a look at the "Programming effort" graph and the "Program length" graph.

That not even an error, it's an usability improvement.

Firstly, usability bugs like this are still bugs.

Secondly, think how that issue surfaced in the first place. Somebody used TensorFlow in their Python code, and their Python code produced unexpected result and/or crashed. Judging by the fix, it is highly unlikely to have produced one of the errors from the above list. As a ML practitioner I can also tell you it likely means somebody had to spend a significant amount of time to get to the cause once they saw the output, because numerical miscalculations are very hard to debug in the first place.

That not a bug that static typing would catch as it is not a type bug.

What do you mean it is not a type bug? You've seen the fix with your own eyes. In the static language that fix would never have been needed, because the consumer would never have made the mistake that caused the miscalculation in the first place. Of course it is a type bug!

This is along the lines of

  def is_valid_email(email): return str(email).contains('@')
  
  class Person:
    ...
    def __str__(self): return $"{self.name} <{self.email}>"
  
  print(is_valid_email(Person(name='Peter @ Work', email=None)))

You can't get a NaN from a None value in Python. (Duck typing still has rules)

The error you have posted isn't a Python error.

Looking a bit deeper, the error occurs in C++ code. C++ is a statically typed language, last time I checked.