When rewards go wrong: reward hacking — Teach a Small AI to Think | unicodesnwrites