Interaction Models (thinkingmachines.ai)

A new research preview has been announced for interaction models, which are designed to facilitate real-time collaboration between humans and AI across multiple modalities—audio, video, and text. This innovation emphasizes the need for AI interactions to be naturally integrated rather than requiring external scaffolding. By employing a multi-stream, micro-turn architecture that processes inputs and generates outputs continuously, these models aim to overcome existing collaboration bottlenecks where users often feel sidelined in automated workflows. The significance of this development lies in its potential to transform human-AI interaction from a turn-based system to a more fluid, real-time engagement, enhancing productivity and user experience. Key capabilities include seamless dialog management, simultaneous speech, and the ability to perform concurrent tasks such as searching or generating user interfaces while interacting. The architecture combines both an interaction model for immediate response and a background model for deeper reasoning, thus marrying responsiveness with intelligence. This approach not only improves how people communicate with AI but also sets the stage for future advancements in multimodal AI applications, marking a significant step forward for the AI/ML community in making collaboration more intuitive and efficient.