Thinking Machines Lab has spent fifteen months raising money, losing key people, and saying very little about what it was actually building. This week, it finally showed something real.
The model, called TML-Interaction-Small, processes audio, video, and text at the same time in short chunks every 200 milliseconds. That means it does not pause to "think" after you finish speaking. It notices a hesitation mid-sentence. It sees what you are doing on camera without you having to describe it. It responds the way a person in the room would. No current product from OpenAI or Google works this way at the same level of responsiveness.
The practical applications are not abstract. A quality inspector on a factory floor could talk through a problem while the AI watches the same conveyor belt. A claims handler could walk through a damaged property on video while the AI flags issues and updates a report in real time. A warehouse manager could ask "alert me if this batch takes longer than the last one" and the AI would actually track time independently, something today's AI tools cannot do without being given a clock to read.
The business case for this kind of AI is real. The question is whether Thinking Machines Lab is the company that will deliver it.
The past twelve months have been rough. Meta approached Murati last year about buying the company outright. She said no. Meta then went and hired five of her founding team members individually, reportedly paying one of them alone more than $1.5 billion over six years. In January, the company's co-founder and chief technology officer was fired after what internal sources described as a misconduct issue, and he immediately resurfaced at OpenAI. Two other founders left with him. The follow-on funding round that would have valued the company at $50 billion never happened.
What remains is a smaller team, a new CTO who helped build PyTorch, a $12 billion valuation from the original 2025 raise, and a genuine technical idea that the bigger labs have not fully matched yet.
The bigger labs will match it. Amazon is already working on similar real-time voice systems. OpenAI and Google are clearly aware of the problem. The window for Thinking Machines to establish itself as the default tool for real-time human-AI interaction is probably measured in months, not years.
For non-technical professionals watching this, there is a useful pattern here. Every large AI lab is currently racing to make AI that replaces human decisions entirely. Thinking Machines is making a deliberate bet in the other direction: AI that amplifies what the human in the room is already doing, watching and listening alongside them rather than replacing them. Whether that philosophical position is commercially smart or not, it is the direction that most traditional industries will actually be able to use first. Heavily regulated sectors, industries where human judgment is legally required, and operations where something can go badly wrong in a matter of seconds are not ready for fully autonomous AI. But an AI that watches alongside a person and flags things in real time? That is useful today.
The interaction model is still a research preview. It is not available to businesses yet. A wider release is expected later in 2026. The benchmarks the company has published have not been independently verified under real-world conditions. These are not minor caveats. A demo in a controlled lab and a working product inside a steel plant or an insurance back office are very different things.
Murati is clearly trying to prove the company still has something to say after a genuinely turbulent year. The technology suggests she might be right. Whether the company survives long enough to make it available to the businesses that would benefit most from it is a different question entirely.