Element-wise arithmetic and comparisons between arrays, scalar promotion, and safe handling of div-by-zero in floats, tied to loss-style numerics and precision pitfalls when subtracting large arrays. Then NumPy universal functions: sin, exp, log, maximum, out parameters to avoid extra allocations, dtype stability, and when to use frompyfunc or vectorize. Together this covers the algebra you write in training loops and the ufunc layer that maps to activations in neural nets and log-space statistics.
