Greg Wilson - software carpentry

- selling Python to scientists

- most computational science is not reproducible, and of unknown quality
- most scientists don't care, because journals and tenure reviewers don't

- improve productivity or quality?
	- RUP: prevent bugs through overdesign
	- XP: catch bugs immediately after creating them
	
- can't sell Python by saying that it makes programming easy
	- have to offer solutions to problems, which happen to use Python

- what are the problems of scientists?
	- they don't know how to program (although they think they do)
	- they don't know how to develop (or the difference)
	- lots of data to manage (in weird formats)
	- many tasks to automate (perhaps involving odd hardware)
	
- best time to teach scientists is in grad school
	- show them how to handle a small software project
	- version control (market as syncing tool)
	- no traction on Python build tools
		- make, ant, IDE builders are too well established
	- data crunching (Python shines on RE, XML, DB)
	- no traction on rapid prototyping (hard to fight MATLAB)
	- GUI easier than Swing, but hard to integrate with legacy code
		- problem with multi-language debugging
		- desktop GUIs will be dominated by browser GUIs
	- testing good
		- but integration is hard
		- no one knows how to unit-test floating point code
		- test I/O, error handling
	- teach design: class design, modularization, debugging
		- in Python it hurts much less than in F90, C++, Java, C#
	- teach process: bug tracking, continuous integration
		- Trac, drproject
		- APIs cool... pipe simulation output into wiki
	- the final frontier: CS 101
		- Python better than C++, Java
		- MATLAB ok...

- open source, free

- www.third-bit.com