Meta has published an updated Advanced AI Scaling Framework detailing how it evaluates and manages risks for its most capable AI models. The framework broadens risk categories to include chemical/biological threats, cybersecurity, and loss-of-control scenarios. An upcoming Safety & Preparedness Report for Muse Spark will document evaluation results and deployment decisions. A key shift in approach: rather than training models on specific rules, Muse Spark is trained on the reasoning behind safety guidelines, enabling it to handle novel situations more robustly. Evaluations are multilayered, covering pre- and post-safeguard testing, live traffic monitoring, and ideological balance checks.
Sort: