The following are the key strengths of TensorFlow:

• Flexible: This Flexibility is coming from TensorFlow's nature of being interpreted. Also, you can see it in the name that it uses a data flow programming mode. The way TensorFlow works, you give the program a graph then it tries to find a node in that graph that's ready to execute, subsequently, it fetches this node and runs it. After this sequence of operations, another set of nodes in the graph become ready to be executed. So TensorFlow does the same with them, fetches them and then executes them. This process of getting the nodes that are ready to be executed and running them is called the Interpreter Loop.
• Expressive: TensorFlow is dynamic as it has some pythonic relations. So you have your full expressiveness and freedom to define your own graph and you are not constrained to anything that can limit you from doing so. Also, TensorFlow is stateful it's similar to the way you program in other environments as you get to use variables.
• Extensible: One of the great advantages of TensorFlow is that it's blackbox modular. So you can come up with totally new data flow operations and attach it to TensorFlow with having the integration headache.

So after mentioning three of the key strengths of TensorFlow, how can TensorFlow team keep these strengths and add more speed by using the just-in-time (JIT) compilation via XLA?