@DKokotajlo makes an obvious point that gets made too rarely: you can have more than one independent AI evaluator. And you should, to balance technical expertise (e.g., human failures vs cybersecurity vs model capabilities), ideology, perspective, etc. (Every grant panel I was