Every AI conversation you've ever had has been a polite fiction of turn-taking, and Mira Murati's new company just said the quiet part out loud.
The Summary
- Thinking Machines is building AI models that process input and generate responses simultaneously, breaking the rigid turn-taking pattern that defines every chatbot today
- Murati is calling these "Interaction Models", a new category distinct from language models or reasoning models
- The goal: make AI feel like a phone call instead of a text chain
The Signal
The current state of AI conversation is weirdly unnatural when you think about it. You type or speak your entire thought, hit enter, and wait while the model processes everything before responding. Every AI model works this way. It's closer to email than conversation. Real humans don't work like that. We interrupt, finish each other's sentences, adjust mid-thought based on facial cues and tone shifts.
Thinking Machines, Murati's post-OpenAI venture, is attacking this limitation head-on. The technical challenge is substantial: models need to maintain coherent output generation while simultaneously processing new input, adjusting context, and potentially revising their response trajectory in real time. This isn't just a UX tweak, it's a fundamental architecture shift.
"You talk, it listens. It responds, you listen. Thinking Machines is trying to change that."
Why this matters for the agent economy: agents that can only operate in strict turn-taking mode will always feel like software, not collaborators. If your AI assistant has to wait for you to finish your entire request before it can start acting, it's not really an assistant. It's a very smart form with a conversational interface. The gap between "tool you use" and "partner you work with" might come down to this simultaneity problem.
The framing as "Interaction Models" rather than just better chat interfaces signals something bigger. Murati isn't positioning this as an incremental improvement to GPT-style architectures. She's suggesting a new category, the way multimodal models became distinct from text-only systems, or how reasoning models like o1 became a category separate from raw prediction engines.
The Implication
Watch how Thinking Machines approaches the interruption problem specifically. Can their system gracefully handle being cut off mid-sentence? Can it incorporate your interruption into its reasoning without starting over? Those edge cases will determine whether this is a real breakthrough or just lower-latency streaming with better marketing.
For anyone building AI products right now, this should make you think harder about conversation design. The turn-taking assumption is baked into every API, every UX pattern, every user expectation. If Thinking Machines ships something that works, your "conversational" interface is going to feel like a telegraph compared to a telephone.