From OpenAI Powerhouse to Thinking Machines Pioneer
Mira Murati, former OpenAI CTO and one of the key leaders behind ChatGPT, has re-emerged with a bold new venture: Thinking Machines Lab. After playing a central role during OpenAI’s turbulent leadership reshuffle in late 2023 and departing the company in September 2024, Murati founded Thinking Machines in February 2025 with a clear thesis. The way humans interact with AI, she argues, has lagged far behind the raw intelligence of the models themselves. Instead of treating the interface as an afterthought, her startup is built around “interaction models” designed from the ground up for natural collaboration. Despite being only about 15 months old, Thinking Machines has already attracted intense industry attention, including reported talks that value the company at around $50 billion and aggressive talent poaching attempts from major tech rivals. The stakes are high: Murati wants AI to behave less like a tool and more like a colleague.
What Makes Thinking Machines AI Different from a Chatbot
Thinking Machines AI is not positioned as just another conversational AI assistant. The company’s so‑called interaction models are built to work continuously, rather than in rigid turns of “you speak, then it speaks.” Today’s mainstream chatbots typically wait for you to finish typing or talking, process your entire input, then respond in one block. Murati’s team sees this as a “bandwidth bottleneck” that limits how much of your intent, context, and expertise can reach the model. Their system breaks interaction into tiny 200‑millisecond chunks, letting it listen, watch, think, and reply at the same time. One part of the model manages the live conversation while another handles more complex reasoning in the background, mirroring how people can talk while planning their next thought. The ambition is clear: move beyond transactional prompts toward fluid human-AI collaboration that feels more like working with another person than operating a machine.
The 0.4-Second Leap Toward Natural Human-AI Collaboration
The headline figure for the first flagship system, TML-Interaction-Small, is its 0.40‑second response time. On paper, this makes Thinking Machines faster than notable rivals such as Google’s Gemini-3.1-flash-live at 0.57 seconds and OpenAI’s GPT-realtime-2.0 at 1.18 seconds. But the speed matters less as a benchmark race and more as a design statement: natural conversation depends on rhythm. Humans unconsciously notice delays of even a fraction of a second in speech. By targeting that threshold, Murati’s startup wants conversational AI that can interrupt, clarify, and react as quickly as a human teammate. This responsiveness enables interaction patterns that traditional chatbot interfaces struggle with—like quick back-and-forth brainstorming, co-editing, or live feedback in meetings. In effect, the 0.4-second response is not just about being first to reply, but about making AI feel present in the moment, not lagging behind the flow of human thought.
From Real-Time Video to ‘Her’-Like Experiences
Thinking Machines’ demos hint at an AI collaborator that is always on, always aware. In one scenario, the model watches a video feed to count exercise repetitions while holding a spoken conversation. In another, it translates speech in real time, and in yet another, it notices when a user slouches—all without pausing the dialogue. This multimodal, continuous awareness moves conversational AI beyond text boxes and single-shot voice queries. It begins to approach the fluid, ambient interactions popularized by the film “Her,” where AI shares the same environment and timeline as its human counterpart. Instead of issuing discrete commands, users can simply live, talk, and work while the system observes and contributes. That shift—from command interface to shared context—could mark a new phase for human-AI collaboration, especially in settings like coaching, creative work, and live support, where timing and subtle cues matter as much as raw intelligence.
Still a Research Preview, But a Clear Signal to the Industry
For now, Thinking Machines AI remains in research preview. The company plans to roll out limited access to research partners in the coming months, with a broader public release expected later this year. Even without general availability, the announcement sends a strong signal about where next-generation conversational AI might be headed: away from static chat windows and toward continuous, shared activity streams. It also underscores intensifying competition for talent and ideas, illustrated by Meta’s reported attempt to acquire the startup and its subsequent hiring of seven founding members, followed by Murati recruiting PyTorch creator Soumith Chintala as CTO. As large players race to refine their own real-time models, Thinking Machines positions itself as the company most willing to rebuild the interaction layer from scratch. If it succeeds, the standard for human-AI collaboration may shift from “smart answers on demand” to “a partner that works alongside you in real time.”