Mira Murati’s Thinking Machines Lab Presents Its First Model
I originally posted this story on Medium.
image by HungryMinded using GPT-Image-2
Currently, most interactions with AI happen on a turn-by-turn basis: the user prompts the model, the model responds, and then the cycle repeats.
illustration prepeared by my Hermes agent
But this is not how humans interact with the world, or with each other.
- We interrupt when it makes sense.
- We keep track of other thoughts while we speak.
- We stay aware of what is happening around us during a conversation.
- We can keep track of time
And this is exactly the stuff that Thinking Machines addresses with their new approach. They have trained a model that has the ability to keep track of things while the conversation keeps going.
This simple demo will give you an idea of how that works:
But this is a super basic example. The model can also handle video inputs and handle several languages at the same time.
Here's another example:
Now that you have a slight idea of what the model does, let's dig deeper.
What is Thinking Machines Lab?
If you’ve never heard of the company behind this model, you are not alone.
Most of the AI talk is concentrated around OpenAI and Anthropic, with some space for Google, Xai, and a couple of Chinese companies.
Apart from this Thinking Machines Lab is still a very new AI company, founded by Mira Murati in 2025. Before this, Mira was the CTO of OpenAI, where she worked on products like ChatGPT, DALL·E, and Sora, and briefly became interim CEO during the whole Sam Altman drama in 2023. When OpenAI’s board fired Sam Altman, then brought him back a few days later.
Mira eventually left OpenAI in 2024 and started Thinking Machines Lab soon after. The company immediately got a lot of attention, partly because of Mira’s track record, and partly because the founding team included several major OpenAI alumni.
It also raised a massive $2 billion seed round, reportedly valuing the company at around $12 billion, which is pretty wild for such a young startup.
Their first public product was Tinker, an API for fine-tuning open-weight language models without having to manage all the distributed training infrastructure yourself.
But this new interaction model feels like the thing that could really put Thinking Machines Lab on the map.
That brings us back to the model.
The Model: TML-Interaction-Small
As we already discussed, the main idea behind this new approach is to bring the way we interact with AI closer to the way we interact with each other and with the world around us.
But there is another issue interaction models are trying to solve: keeping humans in the loop.
With today’s turn-based setup, you usually give the model a prompt, wait for the answer, and then correct it if it goes wrong. But by that point, the model has already moved in a certain direction.
This can be especially painful when the task takes a while to complete, the result is not what you expected, and you have just spent a bunch of precious tokens getting there.
Interaction models make the feedback loop much tighter. You can guide, interrupt, clarify, or redirect the model while the interaction is still happening.
Now, let’s look at how this particular model actually works.
How It Works
The key idea behind TML-Interaction-Small is something Thinking Machines calls time-aligned micro-turns.
Which sounds a bit more complicated than it actually is.
The idea is that instead of waiting for the user to fully finish their turn, the model breaks the interaction into tiny chunks of time. Around 200 milliseconds each.
So, funnily enough, even though they are trying to move away from the classic turn-based structure, they still use “turns” to explain how the model works.
But these are not the turns we are used to.
It is not:
User says something -> Model responds -> User says something else.
Instead, these micro-turns are happening all the time in the background.
Every tiny moment, the model can check what is going on.
illustration prepeared by my Hermes agent
And that is where the interesting part starts.
Because the model is not just sitting there waiting for a finished prompt. It is following the interaction while it is happening.
This is what makes the whole thing feel less like chatting with a text box, and more like interacting with something that is actually paying attention.
- It can listen while the conversation keeps going.
- It can react to visual changes.
- It can keep track of time.
- It can deal with interruptions.
- It can handle audio, video, and text at the same time.
The important part is that the model can stay present while the interaction is unfolding.
You can find more examples and technical descriptions here:
Interaction Models: A Scalable Approach to Human-AI Collaboration
Limitations
As this is the first model of its kind, it still has plenty of limitations.
The first one is easy to notice from the demos: it does not feel perfectly real-time yet.
There is still a short lag before it responds. And this becomes even more noticeable when visual information is involved. Which makes sense. Listening is already hard. But watching a video, understanding what is happening, and responding at the right moment is a much harder problem.
Longer sessions are also still tricky.
Because the model is constantly taking in audio, video, and text, the context window can fill up very quickly. Every second adds more information the model might need to keep track of.
So this is basically the long-context problem, but on hard mode.
Thinking Machines also mentions connectivity as a limitation. For this kind of model to feel good, audio and video need to arrive on time. If frames are delayed, the experience can break pretty quickly.
We will see more on how all of this unfolds when we actually get our hands on a public release.
Closing Thoughts
While this model is the first of its kind, I’ve already seen it eliminate some of the pain points of using voice models today.
The ability to keep track of time is extremely useful for a lot of real-world applications.
I can immediately see myself using something like this to set a timer while cooking and discussing ingredients or recipes at the same time. Or setting a timer for exercises during workouts, while the model guides me through a program, or even provides feedback on my posture.
But the more exciting part is the future of models like these.
I can imagine interaction models combined with something like smart glasses unlocking an experience that feels like a tutorial mode for the real world. The model could see what I see and guide me through tasks of various complexity.
This could enable humanity to achieve so much. From trivial stuff like doing simple repairs on appliances, to more complex things like learning a completely new skill, or even helping someone perform emergency first aid while waiting for professionals to arrive.
As we've seen before. AI progresses fast. And what you saw from today's demos will only get better with time.
