toheedak333

concerning that the models are becoming much better at evading monitors by controlling their visible reasoning, even in deliberate sandabagging (its 25% attempts only). We need more faithful and generalizable monitoring methods.