> They provide empirical evidence that doing the obvious thing - training models not to have "bad thoughts" - causes them to learn obfuscated reward hacking ins

> They provide empirical evidence that doing the obvious thing - training models not to have "bad thoughts" - causes them to learn obfuscated reward hacking instead Just like real humans.