I think the readability issue of vectorized numpy if often ignored (while readability is one of the main strength of python in my opinion). I could not resists to write the example of section 2.2 in Julia:

     using Random, BenchmarkTools
     seq = rand(0:2,10_000); sub = rand(0:2,4);
     # translation of the "readable" but slow python code 
     function_1(seq, sub) = [i for i in 1:(length(seq)-length(sub)) if view(seq,i:i+length(sub)-1) == sub];
     @btime function_1(seq, sub);
     #  93.858 μs (5 allocations: 1.98 KiB)

Which is more than twice as fast as the vectorized (and quite unreadable) python

     %timeit function_2(seq, sub) 
     215 µs ± 1.94 µs per loop (mean ± std. dev. of 7 runs, 1000 loops each)

Having a JIT (like julia, numba...) has really a bit potential to improve code readability and reduce the need of vectorized numpy operations. In some cases, it can even be faster due to be better memory utilization and by computing more directly what you need.