Without AGI there are still cases when the lower probability prediction will be better, and will lead to escaping a local minima. I'd argue that the potential benefits of calibrating that axis dynamically exist with or without AGI.