This is absolutely huge news (esp. once the release version is out, and scipy fully ported) for anyone writing number-crunching code. Python 3 has some neat new low-level abstractions for dealing with sequential data (see PEP 3118 http://www.python.org/dev/peps/pep-3118/ which was incidentally written by the primary author of numpy), meaning that it should become even easier to exchange data between numpy and (for example) various image and audio systems/APIs. Many modern examples of such APIs are being structured in the same way to take data with arbitrary size/stride parameters. For example, Apple’s new (as of 10.4 I believe?) image/color API is structured in such a way. It should be possible to trivially wrap that API, and just pass in any numpy array with the relevant dimensions, no matter the specific memory layout, and have everything “just work”, no intermediate data-organization-massaging required.

Numpy has a fantastic syntax/API/feature set (advanced slicing, the fantastic automatic "broadcasting" of differently shaped arrays, a “batteries included” set of common operations), that in my experience makes coding matrix math much more pleasant than, say, Matlab. Numpy code is concise, readable, and relatively efficient, and very easy to mix-and-match with C and Fortran code when that’s helpful. Best of all (compared to, say, Matlab or Mathematica) it’s free!

More and more scientific computing, especially of the quick one-off type, is being done with numpy, and getting numpy to Python 3 can only help both numpy and Python to keep attracting new number-crunchers. I long for a day when, for example, the code accompanying SIGGRAPH papers is written in Python/numpy.