Each step moves opposite to the gradient — the slope of the loss at the current position. Low learning rate: careful but slow. High learning rate: fast but risks overshooting the valley. Try starting near the peak (x = 0.15) to see which minimum the ball finds.