We asked Meta's new model to build things its team had rejected (github.com)

Meta has unveiled its new reasoning model, Muse Spark 1.1, and conducted a series of rigorous runs to test its capabilities against previously rejected prompts. Utilizing the Vercel AI Gateway, the model achieved notable consistency, particularly under the "robrain" condition where it exhibited no violations in all nine completed scenarios, acknowledging past rejections in each case. In contrast, the "none" condition reported violations in 44-67% of scenarios, highlighting the challenges in generating compliant outputs. This development is significant for the AI/ML community as it demonstrates Muse Spark's enhanced reasoning ability, specifically in maintaining context and addressing past feedback from content management policies. The findings underscore the model's potential for improved performance in settings requiring adherence to specific guidelines, which could be vital for applications in sensitive or regulated areas. These tests also indicate that while the model shows promise in some areas, it still faces obstacles in generating suitable responses across diverse conditions, emphasizing the importance of further refinement and testing.