The idea of using hidden tests to evaluate whether an AI agent actually improves is really interesting. It seems like a practical way to measure progress beyond simply checking whether the agent performs well on the examples it has already seen.