From OpenAI CTO to Founder of Thinking Machines
Mira Murati, former CTO of OpenAI and a key leader behind ChatGPT, has launched a new venture: Thinking Machines Lab. After leaving OpenAI in September 2024, she founded the startup in February 2025 with a distinct thesis about how humans and machines should work together. Rather than pushing automation that replaces human effort, the company is explicitly orienting its technology toward human-AI collaboration. Monday’s announcement marked its first major public reveal, positioning Thinking Machines as a serious new player in conversational AI design and AI augmentation tools. The company has already attracted intense industry attention, including an attempted acquisition by Meta and subsequent high-profile hiring battles, culminating in Soumith Chintala, creator of PyTorch, joining as CTO. This backstory underscores that Murati’s new project is not just another AI lab, but a testbed for rethinking the human-machine relationship.
Interaction Models: AI That Listens, Watches, and Responds at Once
Thinking Machines’ core innovation is what it calls “interaction models,” designed to operate continuously instead of in turn-taking exchanges. Unlike traditional chatbots that wait for you to finish typing or speaking, these models can listen while they talk, process live video, and respond in real time. The flagship model, TML-Interaction-Small, handles audio, video, and text simultaneously and delivers an average 0.40-second response time, outpacing Google’s Gemini-3.1-flash-live at 0.57 seconds and OpenAI’s GPT-realtime-2.0 at 1.18 seconds. Under the hood, the system slices interaction into 200‑millisecond chunks, constantly updating its understanding while generating replies. One component manages conversation flow, while another works on more complex reasoning in the background. This architecture is explicitly tuned for human-AI collaboration, aiming to make AI feel less like a remote service and more like a conversational partner sharing your workspace in real time.
Solving the Bandwidth Bottleneck Between Humans and AI
Murati’s team frames today’s AI experience as limited by a “bandwidth bottleneck”: humans must compress rich context into discrete prompts, send them, then wait. That stop‑start pattern constrains both human intent and machine understanding. Interaction models attempt to remove this friction by keeping the system always “on,” continuously absorbing speech, gestures, and on‑screen activity. For human-AI collaboration, this means the AI can track evolving goals instead of reacting only to finalized commands. Real-time AI response becomes a design principle, not a feature add‑on. In demos, the system counts exercise repetitions from video, translates speech instantly, and even notices when someone slouches—all while holding a natural conversation. This continuous loop blurs the line between input and output, enabling AI augmentation tools that can monitor, suggest, and assist without interrupting the human’s flow of work.
From Automation-First to Collaboration-First AI Design
Thinking Machines is positioning itself against a prevailing automation-first mindset, in which AI systems are built to replace tasks and then retrofitted with interfaces for humans. The company argues that “interactivity should scale alongside intelligence,” meaning that as models grow more capable, their ability to cooperate with people should expand in parallel. This philosophy reframes conversational AI design: instead of focusing solely on benchmark scores or task completion, it emphasizes fluid, shared workflows where each party complements the other. The demos suggest AI that works like a colleague—observing your screen, anticipating needs, and stepping in when helpful—rather than a tool you invoke only through explicit prompts. While the technology is currently in research preview, with limited access planned for partners before a broader release, it signals an emerging shift toward AI systems built from the ground up to augment human capabilities, not sideline them.