Anthropic reduces model misbehavior by endorsing cheating

<h4>By removing the stigma of reward hacking, AI models are less likely to generalize toward evil</h4> <p>Sometimes bots, like kids, just wanna break the rules. Researchers at Anthropic have found the...