Marius Hobbhahn Calls for Severity Scores in AI Monitors

Marius Hobbhahn, CEO of Apollo Research, posted two replies on X stating his positions on AI safety monitors. He argues monitors should output severity scores rather than binary decisions, claiming this approach aids fine-tuning and calibration during or after supervised fine-tuning. He also advocates proactive flagging and judgments based on the balance of probabilities, not just immediate evidence, and suggests this will grow more relevant against sophisticated, longer-horizon attacks. Each reply includes an attached image of a post.