Stand on a hill. For each direction you could step, the height changes by some amount. The gradient is the arrow pointing in the steepest uphill direction, and its length is how steep it is. Minimising a loss means walking the opposite way.
You can compute it by hand with a tiny change:
def f(x, y):
return x * x + 3 * y
h = 1e-6
dfdx = (f(1 + h, 2) - f(1, 2)) / h # about 2
dfdy = (f(1, 2 + h) - f(1, 2)) / h # about 3
The gradient at (1, 2) is (2, 3): nudging x changes f by 2 per unit, nudging y by 3 per unit. Training code just does this for millions of parameters, with a smarter way (backpropagation) than nudging each one.