Sutton might have said you just need a loss function which penalises variance and the model will learn to reduce variance itself. He thinks this will be more ef

Sutton might have said you just need a loss function which penalises variance and the model will learn to reduce variance itself. He thinks this will be more effective than hand coded guardrails. He's probably right. I don't know how you write that loss function mind you. Sounds tricky. But I doubt Sutton was saying it's easy, just that if you can do it then it's effective.