How training-gamers might function (and win) — AI Alignment Forum