Interaction, not intelligence, is the bet at Mira Murati’s new venture. Thinking Machines is positioning its core product around what it calls interaction models, systems that stay alive across a session, track context, and react to user behavior in near real time.
This is a quiet rejection of the prompt-and-response habit that defined the first wave of chatbots, because interaction models treat each exchange as just one event in a longer Markov chain, with state preserved in memory stores and policy layers rather than thrown away after a single reply. Short requests become triggers; longer tasks unfold as sequences of tool calls, database lookups, and interface changes that the model orchestrates while keeping a running internal log.
The company is framing this as infrastructure, not a demo toy. Under the hood sit components that sound closer to reinforcement learning and control systems than to static text prediction, with reward signals tied to user actions, drop-off rates, and task completion instead of generic thumbs-up feedback. In practice that means an interaction model can adjust its own pacing, ask for clarification when telemetry shows confusion, or hand off to specialized submodels when it detects domain specific jargon.
Skeptics will say this is just rebranded agents, yet the emphasis on persistent state and event streams hints at something more operational, closer to a software bus than a chat window. If Thinking Machines can make that bus reliable at scale, the company will have turned the vague idea of conversational AI into a concrete interaction layer for every screen and workflow.
