What Is Thinking Machines Lab?In May 2026, a research team published a paper with the potential to fundamentally change how we interact with AI. The title: "Interaction Models: A Scalable Approach to Human-AI Collaboration." At first glance it reads like a technical report, but at its core it addresses a very practical question that touches all of us: how should we actually work alongside AI?To understand this paper, it helps to know a little about the team behind it and what they set out to achieve, because their mission shapes every technical decision they made.Thinking Machines Lab was founded in February 2025 by Mira Murati, the former Chief Technology Officer of OpenAI. Murati had overseen the development of GPT-4 and ChatGPT, making her one of the most influential figures in the AI industry. When she left OpenAI in September 2024, saying she wanted to "create the time and space to do my own exploration," it sent shockwaves through the industry.The founding team read like an AI industry hall of fame. John Schulman, an OpenAI co-founder, joined as Chief Scientist. Barret Zoph, former VP of Research at OpenAI, came on as CTO. Lilian Weng, Andrew Tulloch, and Luke Metz were also among the founders. Roughly two-thirds of the team came from OpenAI, with the rest arriving from Meta, Google DeepMind, and Mistral.Within little more than a year, however, a wave of departures hit the company. In January 2026, Zoph and Metz returned to OpenAI. Tulloch headed to Meta, and Devendra Chaplot left for xAI. Zoph's exit was especially dramatic: Murati dismissed him citing "unethical conduct," and within the hour OpenAI announced his return, making headlines across the industry.The current leadership stands as follows: Murati remains CEO, John Schulman continues as Chief Scientist, and Soumith Chintala serves as CTO. Chintala is the creator of PyTorch, the open-source machine learning framework that virtually every AI researcher in the world uses on a daily basis — he is, in a sense, the person who built the tools that build AI. Having him steer the technical direction of this research lends it considerable credibility. The team has grown to over 150 people, more than quadrupling since launch, and the company has raised 2 billion dollars across three funding rounds from ten investors.Why Does AI Want to Work Alone?The dominant trend in AI development today is the pursuit of autonomy: the idea that the most capable model is one that can handle long, complex tasks without any human involvement. This assumption runs deep across the industry.But think about how real work actually gets done. Rarely can you define every requirement upfront and walk away. Client demands shift mid-project. You notice a design flaw while writing code. You change direction halfway through a document. Human work is an ongoing cycle of action, reflection, and adjustment. If AI systems leave no room for humans to step in and redirect things along the way, then no matter how intelligent the model becomes, it will fall short in real collaborative settings.The research team calls this the "collaboration bottleneck." Their starting premise is straightforward: instead of asking humans to adapt to the way AI interfaces work, AI should adapt to the way humans naturally work. That shift in perspective is the foundation of everything that follows.The Problem with Turn-Based AI: Like Exchanging LettersAlmost every AI model in use today operates on a turn-based system. You type something, the AI responds, you type again, it responds again. Back and forth, one turn at a time.The core problem is that AI in this model has no sense of the present moment. While you are typing, the AI perceives nothing. While it is generating a response, it receives no new information from you. It is, in effect, like exchanging letters by post.Nobody tries to close an important business deal through letters. Human communication depends on real-time signals: nods, facial expressions, tone of voice, the timing of a pause. These are the things that make conversation feel alive and responsive.The same limitation applies to AI. Jumping in when a conversation is going in circles, catching a mistake the moment it is made, reacting to something visible in the camera — none of this is possible within a turn-based structure.To work around this limitation, existing systems have relied on what the team calls a "harness": external components stitched together to simulate responsiveness. Voice Activity Detection, or VAD, is one such component. It listens to audio patterns to guess when a user has finished speaking and sends a signal to the model accordingly. But VAD is far less sophisticated than the model it serves. It cannot read context or intent. It only detects patterns of sound. The result is a system that sometimes interrupts while you are still thinking, and sometimes misses the moment entirely.The team's argument is simple: interactivity should not be bolted on afterward. It should be built into the model from the start. A model that grows more intelligent should simultaneously become a more natural conversational partner. Responsiveness and intelligence must scale together.The 200-Millisecond Revolution: Micro-Turn DesignThe heart of the proposed interaction model is something called the micro-turn.Where conventional models wait for a complete utterance before beginning to respond, the interaction model processes input and generates output simultaneously in 200-millisecond chunks. To put that in perspective, 200 milliseconds is shorter than the time it takes to move from one sound to the next when speaking — it sits right at the threshold of human perception.What makes this design so significant is that silence, overlapping speech, and interruption all become part of the model's context, rather than edge cases it cannot handle. The model no longer needs to wait for a clean turn boundary. Everything flows as a continuous stream of information.A few concrete examples bring this to life.Real-time simultaneous interpretation becomes possible. While a speaker continues talking in Spanish, the model can deliver the English translation in parallel, without waiting for a pause. In a turn-based system, the interpreter always lags behind. For actual meetings and negotiations, that delay makes it impractical.Consider asking an AI to count your push-ups while you work out. A turn-based model would stay silent until you asked, "How many was that?" An interaction model watches the camera feed continuously and counts aloud in real time: one, two, three.Ask the model to guide a breathing exercise by prompting you to inhale and exhale every four seconds. Because the model has a built-in sense of elapsed time, it can deliver those cues at precisely the right moments, without waiting for any input from you.Two Layers: Balancing Speed and DepthOf course, not everything can be handled in 0.2 seconds. Deep reasoning, web searches, and generating long-form content all take time that a real-time constraint cannot accommodate.The solution is a two-layer architecture. A front-facing interaction model handles all real-time conversation. When a task requires heavier processing, it delegates to a background model running asynchronously. Crucially, the interaction model remains present throughout. It continues to take in new questions, hold the thread of conversation, and weave in results from the background model at a natural moment — rather than delivering them as an abrupt interruption.The analogy to a well-functioning team is apt. A customer-facing representative keeps the conversation going while a specialist works on the answer in the back. When the answer is ready, it arrives smoothly, without breaking the flow.The technical choices supporting this design are equally deliberate. Rather than routing audio and video through large standalone encoder models, the team opted for lightweight processing: audio converted to a format called dMel and handled through a small embedding layer, video broken into 40-by-40-pixel patches processed by a compact architecture called hMLP. All components are trained together from scratch with the main transformer, producing genuinely integrated perception rather than a patchwork of separately trained parts.To address the latency challenges of processing 200-millisecond chunks at high frequency, the team developed a custom streaming session mechanism that maintains a persistent sequence in GPU memory, avoiding costly reallocations with every new chunk. This work has been contributed back to SGLang, an open-source inference framework used across the industry.Benchmarks: A Different Category of CapabilityThe results speak for themselves.On FD-bench v1.5, which measures conversational quality, the team's model TML-Interaction-Small scored 77.8. GPT Realtime 2.0 scored 46.8. Gemini scored 45.5. That is a gap of more than 30 points. On response latency, TML-Interaction-Small clocked in at 0.40 seconds, compared to 1.18 seconds for GPT Realtime 2.0 and 0.94 seconds for Gemini.More revealing are the results on capabilities that existing benchmarks were never designed to measure.On TimeSpeak, which tests whether a model can initiate speech at user-specified moments, TML-Interaction-Small scored 64.7. GPT Realtime 2.0 scored 4.3. On CueSpeak, which measures whether the model responds to verbal cues at precisely the right moment, the scores were 81.7 versus 2.9. On RepCount-A, a visual task requiring the model to count repeated physical actions in a video stream, the scores were 35.4 versus 1.3. On Charades, which tests whether the model can detect when a specific action begins and ends in live video, the scores were 32.4 versus 0.That zero is worth sitting with. Existing models are not merely worse at these tasks. They are constitutionally incapable of performing them. This is not a difference in how smart the model is. It is a difference in how the model was designed.Possibilities and Risks Worth Taking SeriouslyThe potential applications are substantial.In healthcare, a physician conducting an examination could have an AI model watching and listening in real time, surfacing observations like "this patient's breathing pattern has changed since yesterday" or "the dosage you mentioned is 1.5 times the standard amount" — all without the doctor having to stop, type a query, and wait for a response. The AI becomes a continuous presence rather than a tool consulted between tasks.In education, a student working through a math problem at the whiteboard could receive immediate correction — "the sign is reversed there" — at the exact moment the error appears, rather than after submitting work and waiting. The cycle of write, submit, wait, read, repeat disappears.But the risks deserve equal attention.Constant camera and microphone access creates privacy exposure of a different order than existing voice assistants. The volume and intimacy of information that could be processed — and potentially retained — is far greater than anything a turn-based system could accumulate.There is also the question of dependency. When an intelligent, always-responsive presence is available at every moment, people tend to outsource more of their thinking over time. In education especially, where the process of working through difficulty is itself part of learning, this risk needs to be taken seriously.Access is another concern. Low-latency delivery requires stable, high-bandwidth connectivity. In regions where network infrastructure is unreliable, these capabilities will simply not work well. Technology that arrives unevenly across populations tends to widen existing gaps rather than close them.Finally, real-time interaction creates new vectors for manipulation. A bad actor could gradually introduce misleading information across a continuous conversation in ways that a turn-based system's natural pause points might make more visible. The team reports using automated red-teaming to improve robustness, but this is an early-stage solution to a problem that will grow more complex as the technology spreads.The Strategic Picture Behind the ResearchThere is a layer to this story that the paper itself does not address directly.The company is small by the standards of its main competitors. That smallness is also a form of freedom. With less invested in existing architectures, the team could afford to abandon the turn-based model entirely and design something from the ground up. The interaction model is, among other things, a product of that structural independence.When GPT-4o's real-time voice feature launched, it generated enormous excitement. But over time, many users reported something subtly off — unnatural pauses, awkward transitions, difficulty keeping up during complex tasks. These were not symptoms of insufficient intelligence. They were symptoms of interactivity that had been added on top of an architecture not originally built for it. This paper identifies that gap precisely and offers a direct answer to it.The decision to contribute the streaming session work to SGLang is also worth noting. Embedding proprietary innovations into shared industry infrastructure reduces adoption friction and accelerates the process by which a technical approach becomes a standard. Soumith Chintala, who built PyTorch — itself a tool that began as an internal Meta project before becoming the global default for AI research — knows this playbook well. The parallel is hard to miss.Publishing a significant research result at this particular moment, having navigated a year of very public departures, also serves a clear purpose. It signals to the broader community: the work continues, and it is producing results that no one else has produced.What Kind of Relationship Do We Want with AI?The deeper question this research raises is not technical. It is about what kind of relationship we want to have with AI in the first place.Do we want AI as a tool we direct, or as a collaborator that works alongside us in real time? The answer shapes everything about how these systems should be built. The interaction model represents the most serious attempt yet to answer: collaborator.The benchmark numbers are striking. But the more important fact is the collection of zeros — the tasks where every other model simply could not engage at all. Capability gaps that large do not close through incremental improvement. They close when the underlying design changes. And the place where that design change is now happening is the interface between humans and AI: closer to daily life than almost any other frontier in technology.Mira Murati said she left to make space for exploration. A year and a half later, that exploration has produced a model that speaks while it listens, watches while it talks, and tracks time as it flows. Whether or not this specific system becomes the standard, the argument it makes — that interactivity must scale alongside intelligence — is one the entire field will eventually have to reckon with.ReferencesThinking Machines Lab, "Interaction Models: A Scalable Approach to Human-AI Collaboration", Thinking Machines Lab: Connectionism, May 2026. https://thinkingmachines.ai/blog/interaction-models/ Kwa, T., West, B., Becker, J., et al. "Measuring AI Ability to Complete Long Tasks." METR, 2025. Sutton, R. "The Bitter Lesson." Incomplete Ideas, 2019. http://www.incompleteideas.net/IncIdeas/BitterLesson.html Clark, H. and Brennan, S. "Grounding in Communication." Perspectives on Socially Shared Cognition, 1991. Hayek, F. A. "The Use of Knowledge in Society." The American Economic Review, 1945. Bai, et al. "dMel: Speech Tokenization made Simple." arXiv:2407.15835, 2024. Lipman, et al. "Flow Matching for Generative Modeling." arXiv:2210.02747, 2022. Goldman, S. "Former OpenAI CTO Mira Murati finally unveils her Thinking Machines Lab startup." Fortune, February 2025. "Thinking Machines Lab announces Soumith Chintala as new CTO." Social Samosa, January 2026. "Meta hires five Thinking Machines Lab founders including a reported $1.5 billion engineer." The Next Web, April 2026. "One-third of Thinking Machines Lab founding team exits amid fierce AI hiring battle." American Bazaar Online, May 2026. Contrary Research, "Thinking Machines Lab: Business Breakdown & Founding Story." 2026.