Meta joins the pack

This post introduces Meta's Muse Wonder model and benchmarks it against Hex's DataBench. It details Muse Wonder's performance, highlighting its strengths in persistence and adaptation to new information, as well as its weaknesses in cost-efficiency and making unsupported claims at higher effort levels. The post also discusses the behavioral aspects of AI agents, such as the importance of accurate initial claims and the impact of 'checking in' before taking actions, and how these are evaluated within DataBench. The results show Muse Wonder performing at a competitive level, challenging existing top models.