Training kids like this reminds me of training AI itself:
Make the reward getting a better grade and the kid will only optimize for that. Eventually the kid (the agent) will be out in the wild and they will have a weak understanding of what the reward is and how to optimize for it (it will be harder for a fine-tuned kid to do domain adaptation towards different and more abstract fields and problems).
Make the reward getting a better grade and the kid will only optimize for that. Eventually the kid (the agent) will be out in the wild and they will have a weak understanding of what the reward is and how to optimize for it (it will be harder for a fine-tuned kid to do domain adaptation towards different and more abstract fields and problems).