Smaller monitored models catch deceptive AI better than large ones

Original: Great work on building better trusted monitors from the latest @MariusHobbhahn x @MATSprogram team! 👁️

Source: x.com ↗