This is a self-fulfilling prophecy.

Someone tries to write an analysis routine in 3.x, but then is quickly told "nobody in science uses 3.x, because of <vague reasons>". Result: everyone sticks to 2.7, so nobody ever even tries to build against 3.x when writing packages.

I switched to python3 after measured a large ETL process converted. As scientific packages are rely on FFI and the actual code is running outside of python I'm puzzled with your comment about speed.

Neither of these assertions are true. I use Py3 to teach graduate-level scientific computing and machine learning with great success, and the scientific Python community is making a very strong push for adoption. All major packages in the scientific stack now support Python 3.

That's interesting, do you have a good source on that? I switched to 3.x for development but I do academic projects too.

I don't know about speed, most of the modules where speed matters, like Numpy, rely on calling C routines somewhere down the line. The version of Python shouldn't make a significant difference.

Numpy, Scipy and Pandas work with Python 3, Jupyter certainly does. That covers probably 80% of scientific grunt work. For specialist applications, OpenCV 3.0 works, as does scikit-learn/image. I can't speak for other development like web though.

I think the main problem is that if you're using someone else's code in academia, it's likely to be written in 2.7 and you would have to go through and update everything that's not back-compatible.

slower.. <citation needed>