I wanted to announce JET, a different take on how to make numeric Python much faster! It's very much inspired by TensorFlow -- a computation graph is compiled to C++. And there are constants, variables and placeholders, like in TensorFlow.<p>But, it's converted to a regular C++ class, and compiled using e.g. GCC instead of being executed with a custom scheduler (as it is in TensorFlow). This makes execution for small programs much, much faster.<p>Compared to NumPy it's up to 20x faster.