For me, the only place where Python truly excels versus the competition is Numpy. The easy python scaffolding gets you to the heart of a problem in a friendly and flexible way, and then you can interface easily to this big V8 of a compute engine (especially if you combine it with the MKL libraries). The ecosystem around Numpy is astronomically good.

I have fallen in love again with Python (2) for this reason. I tried 3 but I don't need all it's accountant-like seriousness. It's less fun to code in. When I need to design for servers or for performance I will do golang or ocaml. When I need simple scaffolding for numerical work, and scripting for prototyping integration of various systems, Python 2.7 is like an old, unfussy, friend, helping me along my long and winding data wrangling road with a minimum of cognitive tax. It just works. Long may it live.

In my view the Python powers should look at numerical / data science and throw all of Python's weight behind that, and forget about the web, where they've already lost.

Thanks for taking the time to read. I agree with most of what you've said, but of course I still don't think Python has lost on the web, though that was my concern that sparked writing this post.

What are your thoughts on the future of Python + NumPy vs. MATLAB or R? It seems the numeric and scientific community is embracing Python more, but what I've heard is mostly anecdotal.

I use R extensively, even more than Python, but it is not a robust language when your problem moves from exploration to anything resembling production. The memory management appears to be the problem. Hugely inefficient and slow even with 3.x. I get crashes frequently, even under Linux and I run production on R. Moreover, Numpy's two orders-of-magnitude speed increase opens up a new class of problem that R just can't handle, especially in the streaming data space. Pandas has made a big improvement in taking Python towards R, though it would be great if it were multidimensional. Where R has Python beat hands down is in its "vectors everywhere" functional philosophy. List comprehensions just don't match the power, and vector indexing is uneven even in Pandas.

I have less experience with Matlab for cost, and lock in reasons. I need to have the option to distribute my code easily.