OpenAI shipped a full-duplex voice system in six months—here's how GPT-Live actually works

GPT-Live ditched the turn detector for continuous streaming inference, stateful model handoffs, and async delegation. The engineering behind sub-second voice responsiveness is wild.