Why add the second term instead of subtracting it?

I think it's because the derivative is negative. We want to adjust the x by x:=x-delta_t * derivative. The two negative signs cancel out.

Isn't this the basis for neural networks?