Section 3 of the note is costs — write the two parameter counters: dense_params(n_in, n_out) = n_inn_out + n_out (weights + biases) and conv_params(kernel, c_in, c_out) = kernelc_in*c_out + c_out. Print the dense 256→256 count, the conv k=3 16→16-channel count, the integer ratio, and the closing line "same conv params at ANY input length - weight sharing pays the bill".