Start at the surface. Descend when you're ready.
How a model learns by following the slope of its own error.