Meta's Muse appears to use an OpenAI model labeled muse-special (mouse.dev)

Meta's exploration of its Muse model has revealed the presence of an intriguing entry labeled "azure/muse-special," suggesting potential integration with OpenAI technologies. Digging into the system's logs, it appears that while the majority of sessions utilize Meta's internal Avocado model, the muse-special subagent seems to employ OpenAI's API, evident from unique signature tags and encrypted payload formats that align with OpenAI’s specifications. This discovery raises questions about the underlying architecture of Muse and the potential for routing requests to various AI models, including those from OpenAI and Anthropic, based on task effectiveness. The significance of this revelation lies in its implications for AI model usage and development within Meta. The architecture allows for flexible A/B testing and model optimization without user intervention, highlighting a shift towards a more dynamic AI environment. While it's unclear if Meta is actively distilling knowledge from other models, the presence of multiple API keys and client structures suggests they are keen on leveraging the best available technology to enhance Muse’s performance. Overall, this insight into Muse's functionality offers a glimpse into the future of personalized AI agents and underlines Meta's continued commitment to innovating in the AI/ML landscape.