Eight months after leaving OpenAI as its chief technology officer, Mira Murati has broken her silence on what Thinking Machines Lab is actually building—and it represents a philosophical departure from the company she helped shape into a $150 billion juggernaut.
The concept is called "interaction models," and if Murati's framing is to be believed, it could represent the most significant rethinking of how AI systems work since the transformer architecture that enabled ChatGPT in the first place. Rather than simply training models to be smarter, faster, or more capable, Thinking Machines is focused on training models to understand how individual humans actually want to communicate.
The thesis: intelligence isn't the bottleneck
The current AI paradigm treats every conversation as a fresh start. Whether you're a software engineer debugging code or a novelist brainstorming plot points, GPT-4 or Claude approaches you with the same conversational cadence, the same assumptions about what you need. Thinking Machines argues this is backwards. The models are already smart enough for most tasks—what's broken is the interface layer between human intent and machine capability.
Interaction models, as Murati's team describes them, would learn not just what you ask for, but how you ask for it. They would adapt their communication style, their level of detail, their willingness to push back or simply execute. In theory, an interaction model used by a busy executive would behave fundamentally differently than one used by a graduate student, even when answering identical questions.
Why now, and why her
Murati's departure from OpenAI last September was abrupt and unexplained, coming just months after she briefly served as interim CEO during Sam Altman's dramatic firing and reinstatement. She took with her a reputation as one of the few people who understood both the technical and product dimensions of large language models at the highest level.
Thinking Machines has reportedly raised significant capital—though the company has not disclosed figures—and has been quietly recruiting researchers from DeepMind, Anthropic, and her former employer. The interaction models announcement suggests the company is further along than the typical stealth-mode startup, though no timeline for a product release was provided.
The competitive landscape shifts
The announcement arrives at a peculiar moment for the AI industry. OpenAI, Google, and Anthropic are locked in an increasingly expensive arms race to build larger, more capable foundation models. Murati appears to be arguing that this race, while impressive, may be solving the wrong problem. If she's right, the companies pouring billions into raw intelligence gains could find themselves outflanked by a startup focused on something far more mundane: making AI actually pleasant to use.
Our take
Murati is making a contrarian bet that the AI industry has become so obsessed with benchmarks and capabilities that it's forgotten about the humans on the other end of the conversation. It's a compelling thesis, and she may be uniquely positioned to execute on it. But "interaction models" remains frustratingly vague as a technical concept—it could be a genuine paradigm shift or sophisticated marketing for what amounts to better fine-tuning. The proof will be in the product, and Thinking Machines has yet to show one.
