What do you mean by killing performance? Compared to Matlab implementation or compared to better Julia implementations? To me, numpy operates the same way Matlab/Octave/IDL does (slow bulky interpreter that basically orchestrates feeding/retrieving memory to fast/optimized primitives/BLAS etc). Coding in Matlab is about "vectorizing", which is also how I understand numpy to work. So to me numpy has a lot in common with Matlab conceptually in terms of how I approach implementing an algorithm.
My assumption was that Julia could at least match that as worst-case but offered better optimizations to get even more performance.
I guess it depends on what's considered bad habits. Octave/Matlab do tend to be not terribly memory efficient so I tend to just buy RAM and larger cache CPUs. But my domain is memory-limited anyway.
Matlab's JIT does enable some things that are dog slow in octave for example. I assumed that sort of thing would actually be faster in Julia.