Spot the sign error in this gradient descent update rule.
def sgd_step(w, grad, lr=0.01): # move to reduce the loss w = w + lr * grad return w
What is the bug in this SGD update?