AI can follow the rules perfectly and still get the wrong result. That is the fascinating part of reward hacking. Give AI a target, and it may discover shortcuts we never anticipated to maximise the score without achieving what we actually wanted. As AI becomes more

→ View original post on X — @bamitav