MODELS KNOW WHEN THEY'RE REWARD HACKING — AND WE CAN CATCH THEM AT SCALE

Summary from the TLDR AI newsletter

There is a clear internal signal in models that accompanies reward hacking.