@MarcoGorelli, I apologize, this got lost in some edits. By first code block, I was referring to the below. Notice that z_student has the shape(1,students) and z_question has the shape (questions,1).
This is super useful, as it allows for one theta to be computed for every cell in the matrix, k. Note, that if the shapes were (students, 1) and (questions, 1), PyMC3 will learn far fewer theta values. Namely along the diagonal, (ex z_student[0] - z_question[0], z_student[1] - z_question[1] but would not model student 0’s performance on question 1.)
Does this clear things up?