Mira Murati is firing back. Following her high-profile departure from OpenAI and explosive courtroom testimony against Sam Altman, the former CTO just unveiled the exact technology meant to dethrone her former boss. Her new startup, Thinking Machines Lab, revealed a breakthrough AI system that fundamentally breaks the clunky rules of modern chatbots.
It listens while it talks. The company announced a research preview for a new class of interaction models built around a true full duplex mode. This means users can finally interrupt the AI mid-sentence and it will process the incoming audio in near-realtime.
The 0.40-Second Response Latency
The underlying architecture relies on a multi-stream, micro-turn design. The prototype, officially dubbed “TML-Interaction-Small,” is incredibly fast. It boasts a staggering response latency of just 0.40 seconds.
This effectively kills the awkward turn-based pausing that defines current voice assistants. According to a preview of the near-realtime AI, the system processes audio, video, and text inputs concurrently to generate immediate outputs without forcing the user to wait their turn.
Right now, the system remains locked in a restricted research preview. But a breakdown of the real-time interaction models confirms the startup plans to open wider access in the coming months. A broader public release is slated for later this year.
The timing of the reveal is entirely deliberate. The tech world is currently watching Murati’s videotaped testimony in Elon Musk’s $180 billion lawsuit against OpenAI. She publicly accused Altman of chaotic management, giving a testimony detailing her concerns that ultimately catalyzed her exit from the massive firm.
Since leaving, she secured major funding and aggressively poached key researchers directly from OpenAI. This escalating battle for top-tier talent is ripping through the technology sector as rival labs rush to match the speed of these new interaction models.
Why Full Duplex AI Crushes the DeepMind Advantage
The shift to full duplex interaction models forces an industry-wide pivot. Current multimodal models from Google DeepMind and OpenAI simulate conversation by taking turns. You speak, the machine processes, the machine speaks. Murati’s 0.40-second latency destroys that paradigm.
Thinking Machines is leveraging the exact talent that built ChatGPT to eliminate the biggest friction point in consumer AI adoption. The waiting is over. If a machine can truly listen and adapt its response mid-sentence without crashing, the barrier between software processing and human collaboration completely vanishes.