Mira Murati 的 Thinking Machines Lab 推出实时交互模型

thinking machines lab published a research preview of its first model, arguing that current voice ai relies on unintelligent 'harnesses' that rigidly segment conversation. its new 'interaction model' processes audio and video directly in the transformer using a shared 200-millisecond clock, enabling fluid, full-duplex conversation where input and output streams interleave. this design allows the model to react proactively or interrupt without waiting for complete utterances, a feature lacking in competitors like gpt-realtime-2. to maintain real-time speeds while supporting complex reasoning, the system pairs the fast interaction model with a background model that handles asynchronous tasks like tool use and research within the same context.