"There's very little to be gained by reading pure Fortran code, especially if you don't know the mathematical tricks/optimisations used beforehand. The Python code, on the other hand, stays (relatively) readable."

Depends. His first Python code was:

  x = np.asarray(x, dtype=float)
  N = x.shape[0]
  n = np.arange(N)
  k = n.reshape((N, 1))
  M = np.exp(-2j * np.pi * k * n / N)
  return np.dot(M, x)

I can write more or less the same in Fortran:

  n = reshape([(i, i=0,len-1)], shape=[1,len])
  k = transpose(n)
  M = exp(-2 * j * pi * matmul(k,n) / len)
  res = matmul(M, x)

But of course Fortran is a typed language, so I need to explicitly write the types for everything, so my whole function gets longer

  pure function dft_slow(x, len) result(res)
    complex, intent(in) :: x(len)
    integer, intent(in) :: len
    complex             :: n(1,len), k(len,1), M(len, len), res(len)
    integer             :: i

    n = reshape([(i, i=0,len-1)], shape=[1,len])
    k = transpose(n)
    M = exp(-2 * j * pi * matmul(k,n) / len)
    res = matmul(M, x)
  end function dft_slow

and so yes, maybe less readable.

P.S. One must also define somewhere global constants

    complex, parameter  :: j = (0,1)  ! imaginary unit
    real, parameter     :: pi = acos(-1.)

for the above function to work. Also everything is by default in single precision here.