|
I appreciate this article being posted, and have the utmost respect for NumPy developers. The urgency and discrepancy between use of certain important open-source libraries, and their support, is bewildering sometimes. As I was thinking about it, though, I'm not surprised NumPy hasn't been funded before. The reasons why say a lot about biases in memory. It wasn't that long ago that the sorts of things NumPy does were seen as fairly niche, and in the domain of statistics or engineering. It's only with relatively recent interest in AI and DL that this has been seen as within the purview of Silicon Valley-comp sci-type business, as opposed to EE or something different. I still am kind of a little disoriented--the other day, looking through our university's course catalog, I realized that certain topics that would have been taught in the stats or psychology departments are now being seen as the territory of comp sci. Statisticians have written excoriations about being treated as if they don't exist, as comp sci blithely barrels forward, reinventing the wheel. I'm not meaning to take sides with these issues, only pointing out that I think the world we live in was very different not so long ago. It might seem puzzling that NumPy hasn't had more funding, but I think that's in part because what it's most profitably used for now wasn't really seen as much more than academic science fiction not too long ago. The other part of it too, is that until relatively recently, if you were to do numerical heavy lifting, you'd almost certainly be expected to do that in C/C++ or maybe Fortran. There's a tension in numerical computing, between the performance and expressiveness that's needed, and Python is on one end of that continuum, far from the end that is traditionally associated with complex numerical computing. Sure, you had things like MATLAB with Python in the same functional role, but those were largely seen as teaching tools, or something that engineers did for one-off projects, having learned to do that in school (I still think the use of python in ML derives from the use of Python as a teaching tool in uni). I'm not trying to knock Python or NumPy or anything, just kind of trying to convey a different perspective, which is that I can remember a time not too long ago when the use of Python in numerics was seen as primarily didactic in nature, or for limited circumscribed applications. FWIW, it seems to me Python is kind of on a path similar to what happened with javascript, which was treated as kind of an ancillary helper language on the web, until Google started pushing its limits. Then there was browser wars 2.0, and huge efforts put into javascript, and it became a main player in network computing. To me, there's a similar trend with Python: it really kind of existed as a language for prototyping and scripting tasks, and now finds itself in a different role than it has been used for traditionally, and projects in that area are getting an influx of money accordingly. What I see happening is (1) a blossoming diversity of numerical computing communities (Haskell, Python, Julia, Kotlin, Scala, Rust, Go, etc.), due to competition and variation in application scenarios and preferences, (2) a huge influx of resources being put into Python to make it more performant, or (3) people jumping ship from Python into one of those other platforms to get more bang-for-the buck [or (4) some combination of all of these.] |