Jakub Pachocki wrote on Sunday that no lab has tied its models to human goals well enough. He says reading a model's chain of thought to watch it is working less and less. He wants safety thresholds made mandatory and checked by outside auditors or governments.