Towards understanding-based safety evaluations — AI Alignment Forum