The short version
We still talk to AI the way people used to write letters. I write, it answers, I write back. Taking turns, in moves. It is precisely this turn-taking procedure that is the bottleneck, not the intelligence of the model. When the machine keeps thinking while I talk, and I keep thinking while it works, something new emerges. I call it co-thinking. And I believe it is more than a product feature. It is the third discipline of leadership.
Much in this text is conjecture. I write it down in concrete sentences anyway, because only those can be refuted.
Why turn-taking is the real problem
Anyone who works seriously with language models knows the feeling: you spend the whole time building around a constraint. The model does not know what is happening while I am still typing or speaking. It does not see that I pause mid-sentence. And while it is generating an answer itself, it is blind to everything I think in the meantime. Mira Murati put it sharply: while such a model is thinking, it is almost deaf and blind, it perceives nothing of what else is going on, and while I speak, it does not register how I say it.
This is not a cosmetic problem. It narrows the channel between my knowledge, my intention, my judgment and what reaches the model. I have to break my thinking into clean packets, send them off, wait. In a real conversation with a clever person I never do that. There I think along in parallel, interject, correct myself, react to a frown before a word has been spoken. Heinrich von Kleist described this around 1805/06, in his essay on the gradual formation of thoughts while speaking: the thought only comes into being in the act of speaking, addressed to a counterpart who need not be cleverer for it, only present and attentive. That is exactly where the information sits that gets lost in turn-taking mode: in the silence, in the thinking, in the interrupting. Turn-taking is a narrowband channel for something that is really high-bandwidth.
What is interesting is what happens when you lift that constraint even halfway. With today’s tools this is already roughly possible: I simply talk into it, the machine thinks – the more complex the question, the longer – it reports back, we iterate. In the meantime I switch context and do something else. This is not yet co-thinking, it is its precursor, and the context switches are the price of the makeshift. But even the precursor changes how the work feels: the machine does not feel like a teacher explaining something to me, but like a very smart colleague who inspires me. In the target state the price falls away too: the machine works in the background while I keep thinking up front, and both remain a single conversation.
This is no longer theory
The idea itself is older than any language model: J.C.R. Licklider described the man-computer symbiosis back in 1960, joint decision-making without rigid programs, and Doug Engelbart wanted to augment the intellect, not replace it. Six decades later, working at exactly this spot, is Mira Murati, the former CTO of OpenAI, with her Thinking Machines Lab. On 11 May 2026 the lab presented a research preview they call Interaction Models. The core idea: interactivity belongs inside the model itself, not strapped on top as a collection of helper components.
The architecture hits the point almost word for word. The preview describes a time-aware Interaction Model for real-time presence and an asynchronous Background Model for longer reasoning, tool use and planning, both on a continuously shared context. The Interaction Model stays permanently live, takes in audio, video and text in 200-millisecond chunks and reacts immediately. If a task needs deeper thinking, it hands it off to the Background Model – with the full conversation as context, not as an isolated question. The results flow back as soon as they are ready and get woven into the conversation at the moment that fits.
Thinking Machines describes it themselves with an image that is exactly my co-thinking: like a person who keeps you in the conversation while a colleague in the background looks something up and passes the notes forward in real time.
One has to add: this is a research preview, and the technical community argues about how much of it is really new. As proof it does not hold up, as a directional signal it does. For me it is not decisive anyway whether this particular architecture wins. What is decisive is that it addresses an old question in a new technical way: how does human judgment stay present during the work, not only at the end?
At Bloomberg Tech 2026, Murati even rejects the tame image of the human in the loop. It sounds like a checkpoint where you nod things through at the end and then it is fine. She means something else: a tandem. Both pedal, on the hill the stronger one pushes harder, but both hands stay on the handlebars. A system built for collaboration, not for sign-off.
Behind this lies a directional decision. There is a fast path, to build AI as purely autonomous, detached from the messiness of reality and from daily human experience. And there is the neglected path, to bring the machine closer to where people’s knowledge and intention actually sit. Murati names exactly that as her commitment: human agency should grow, not disappear. The lab’s mission statement draws the same line: collaborative general intelligence, an AI that fits the messy way we humans actually work together.
In other words: the bet of the most serious people in the field is no longer just raw model strength. It is the quality of the collaboration itself.
The bold claim: co-thinking makes a whole class of problems solvable
If that is true, something follows that matters to me. A whole range of problems I grind away at today are not knowledge problems at all. They are thinking problems. I cannot get further because I lack a sparring partner who is fast enough, patient enough and broad enough to think with me, rather than to answer me.
My claim: co-thinking can make a whole class of such problems solvable. Not because the model knows the better answer, but because the continuous, parallel loop between me and the machine produces a thinking result that neither of us would have had alone. That is testable, and that is exactly what I want to think through in public, instead of asserting it.
A tool for thinking
Murati boils this down to a term that is bigger than productivity. The most advanced AI systems, she says, are the most incredible tools for thinking that humanity has ever had. They change not only how fast we think, but what we can think about at all.
Her example lands. Imagine you had to multiply with Roman numerals. It was today’s digits that opened up an entire field of mathematics, so far that a child today computes what scholars once labored over. Deep tools – language, writing, numbers – have always shifted what we think about, not just how fast.
That is exactly my bold claim, one step more concrete. Co-thinking is such a tool. It does not deliver the better answer to a question I can already ask. It widens the space of questions I dare to think at all. A class of problems then resolves not because one side is cleverer, but because the joint thinking reaches a thought that previously lay beyond my reach.
What co-thinking changes about alignment
The part most important to me comes at the end, and it was unclear to me for a long time: co-thinking is not only more productive, it is safer. It does not solve alignment. But it addresses an operational layer of alignment problems: specification, oversight and traceability within the running work process.
First, the specification gap. The classic source of error is that I can never fully describe in advance what I want, and the model then optimizes for the wrong thing. In turn-taking mode I have to throw the complete brief out front and hope. In co-thinking I steer continuously, I interject, correct myself mid-thought, add my judgment while the work is taking shape. More of my intention reaches the model at all. The gap between what I mean and what arrives gets smaller.
Second, control – and not as a sign-off stamp at the end, but as continuous co-steering. Both hands stay on the handlebars. I can interrupt a wrong direction before it builds up, instead of discovering the misalignment only in the finished result. This is oversight at the speed of the work, not after the fact, and it is the counter-model to the autonomous agent that runs off and whose errors you only see once they have become expensive.
Third, traceability. A system that keeps passing me the notes lays its thinking open while it thinks. I can check and redirect it, instead of judging a black box at the end. Thinking Machines names exactly this as the core: turn-taking limits how much of a person’s knowledge, intention and judgment reaches the model, and how much of the model’s work can be understood at all.
Murati says it openly: whoever builds this way steers research toward outcomes that are more value-aligned. Alignment then falls out as a byproduct of the collaboration, on top of the usefulness. And she turns the argument into the dimension of time. Whoever takes the human out of the loop now squanders the chance to get it right later, when the systems are even more capable. Keeping the human in the loop today is therefore not a brake, but the precondition for the next stage to become safe at all.
By now there is a name for the flip side: cognitive debt. Early, still cautiously-to-be-read studies suggest that people who outsource their thinking to the machine build up less of their own cognitive activity – and do not simply get it back when the machine is gone. This is not an argument against AI, it is an argument for the right operating mode. Dumbing down is the result of thinking-for-me. A counterpart that thinks with me trains the judgment instead of replacing it. Kleist already knew: the counterpart makes one cleverer, not dumber.
From this follows a point that matters to me socially. There is a difference between an AI thinking with me, for me, or at me. A machine that thinks along, transparent and correctable, keeps me in the driver’s seat of my own judgment. One that decides for me or persuades me takes it from me. The objection is obvious: a system that reads my intention better, because it picks up tone, timing and hesitation, could also influence me better. That is why co-thinking is only safer if transparency, correctability and role clarity are built in. In a world where manipulation via AI and social media is real, co-thinking is therefore not just a comfort argument, but a stance: the human remains the authority that understands and answers for it. The less rupture the next capability jumps create, the more likely we keep our hand on the handlebars.
What co-thinking is for, and what it is not
As convinced as I am of all this: co-thinking is not the answer to everything. When I sort my work with AI, I see four operating modes, and the question that orders them is always the same: where does my judgment sit in time?
The assembly line. Fixed cadence, fixed path. Everything repetitive that follows clear rules belongs here. The judgment is cast into the design once; after that the thing runs by itself, and the human comes in only for exceptions.
The autopilot. Here the path is open, but the goal is fixed. The machine plans, uses tools, sometimes works for hours on its own, and my judgment sits at checkpoints: briefing, interim status, sign-off. The goal need not come fresh from me each time. It can stand as a standing instruction and be triggered by an event, an incoming email for instance. The trigger does not replace the goal, it starts it.
In practice these two operating modes rarely occur alone. The assembly line hands off to the autopilot when a case falls out of the rule, and the autopilot hands back what has become routine again. Today companies orchestrate workflow engines and people for this. Tomorrow they will orchestrate workflow engines, agents and people – the same architecture, one player more. It is exactly this integration we are building at the Exigo AI Studio as a product: digital work as a service, where the human is no longer stuck in every step, but at the places where their judgment is needed.
The correspondence. I ask, it answers, I ask back. This is the mode this text began with, and my bet is: it is a transitional form. It exists only because the technology could not do anything better so far, and it will be dissolved in both directions. The repetitive sinks to the autopilot, the judgment-laden rises to the tandem. No one will mourn the correspondence, as little as the letter, now that we can call each other.
The tandem. Co-thinking. Strategy, scope, judgment. Wherever the problem is not yet cleanly described, where the question itself is part of the work, no autopilot helps me that runs off and eventually delivers a result. There I need the continuous, parallel loop this text is about. And because the term is currently making a career, a demarcation: when consultancies talk about the co-thinker today, they almost always mean the better correspondence. I mean the parallel regime.
This also dissolves a dispute that the alignment debate likes to run as a yes-no question: the human in the loop. In fact it is a question of assignment. On the assembly line the human stands at the exception. With the autopilot they are in the loop in the classic sense, at checkpoints, and there the image is right: the better the checkpoints, the more the machine may do alone. In the tandem, by contrast, there is no my loop and your loop, there is only a shared one. Murati therefore does not reject the checkpoint, but the checkpoint as a universal image.
None of these operating modes is the better one. As early as 1958 the Harvard Business Review described leadership as a continuum, from directing to joint decision-making, and the mastery lay then as now in choosing the right point. It only gets expensive when you confuse them. Whoever rides in the tandem what belongs on the assembly line wastes their attention. Whoever leaves to the autopilot what really needs judgment gets very convincing wrong answers very quickly.
What speaks against it
I take three objections seriously, and they make the case better, not worse.
First, the empirics. A large meta-analysis in Nature Human Behaviour over 106 experiments found in 2024 that human-AI teams on average perform worse than the better of the two alone, especially on decision tasks. That sounds like a refutation. Except: practically all the teams studied worked in correspondence mode. So the study does not measure the limit of joint thinking, but the cost of turn-taking – and thereby rather confirms the bottleneck this text is about. On open creation tasks, by the way, the same analysis found gains.
Second, the chess argument. In chess there was the centaur phase, human plus machine beat the machine alone – it was short, today the human only gets in the way. Does co-thinking end the same way? For everything that behaves like chess: yes. Closed, verifiable tasks migrate sooner or later into the autopilot, this text claims so itself. But strategy, scope and judgment are not chess. There is no engine that checks whether my question was the right one.
Third, scaling. Co-thinking consumes the scarcest thing I have, my attention. It does not scale to a fleet of a hundred agents. True – and that is exactly why the assignment of operating modes is the core competence we come to next: tandem for the few questions that really turn something, autopilot and assembly line for the rest.
Why I call this the third discipline
Leadership has so far played out in two disciplines. The first is leading people: giving direction, distributing responsibility, building trust. The second is leading systems: commissioning processes and increasingly agents and gathering results. It is closer to the first than you might think, by the way: agents need briefings instead of shouts and feedback instead of hope. Whoever can hold an employee review brings more to this discipline than any prompt course. Both still have one thing in common: I delegate tasks.
Co-thinking is different, because I do not delegate a task, I share the thinking. I am not commanding and I am not automating, I am thinking together. That is the third discipline – I call it Third Leadership – and it demands four abilities.
Framing: making unclear problems speakable. I have to endure and voice my half-finished thoughts, instead of first formulating the finished brief.
Steering: intervening during the thinking. I have to manage context switches without losing the thread, and correct a crooked direction while it is still cheap.
Judgment: assessing machine suggestions before they solidify. I have to learn when I am the stronger one on the hill and when the machine is.
Mode choice: cleanly assigning what belongs where – to the assembly line, to the autopilot, into the correspondence, into the tandem, or to a human. This assignment is the core competence that holds the three disciplines together.
Whoever masters this early leads not only themselves and their team, but also a thinking partnership. That is the actual competence of the coming years.
What this means for companies
The central AI question in companies is thus no longer “which model do we have?”, but “where does human judgment need to sit in the workflow?”. For me, and for what we are building at the Exigo AI Studio, this has a clear consequence. AI is not rolled out on top of everything else. It is embedded sharply at certain points, where the joint thinking makes the difference. Not the chatbot for every employee, which in isolation stays astonishingly incompetent, but the orchestration that brings specific knowledge and joint thinking together.
The operating modes also show how companies will buy AI in future. What belongs on the assembly line and in the autopilot no longer has to be procured and operated as a tool by anyone. You can buy the finished result: complete outcomes instead of software, work done as a service, where your own judgment only sits at the agreed checkpoints. That is exactly our product. The tandem, by contrast, cannot be bought and cannot be delegated. Where the joint thinking makes the difference, it is the work of leadership itself, and for that it takes the third discipline.
Good ideas are shared. That is why I write this down openly, while it is fresh.
Sources
- Thinking Machines Lab, “Interaction Models: A Scalable Approach to Human-AI Collaboration”, 11 May 2026 – thinkingmachines.ai/blog/interaction-models
- Mira Murati, mission statement (X, July 2025): “empower humanity through advancing collaborative general intelligence”
- Bloomberg Tech 2026 (4 June 2026), Murati in conversation with Emily Chang, “Thinking Machines’ Murati on AI’s Next Chapter” – youtube.com/watch?v=A_jIpryR5js
- MarkTechPost, technical analysis of the Interaction Models, 13 May 2026
- Heinrich von Kleist, “Über die allmähliche Verfertigung der Gedanken beim Reden” (On the Gradual Formation of Thoughts While Speaking), written around 1805/06, published posthumously
- Robert Tannenbaum / Warren H. Schmidt, “How to Choose a Leadership Pattern”, Harvard Business Review, 1958
- J.C.R. Licklider, “Man-Computer Symbiosis”, IRE Transactions on Human Factors in Electronics, 1960
- Vaccaro/Almaatouq/Malone, “When combinations of humans and AI are useful”, Nature Human Behaviour, 2024 (meta-analysis over 106 experiments)
- Ethan Mollick, “Centaurs and Cyborgs on the Jagged Frontier”, One Useful Thing, September 2023
- Sean Goedecke, critical analysis of the Interaction Models, seangoedecke.com, May 2026
- Amelia Wattenberger, “Why Chatbots Are Not the Future”, May 2023
- Kosmyna et al. (MIT Media Lab), “Your Brain on ChatGPT” – EEG study, coins the term “Cognitive Debt”, 2025 (preprint, n=54, cite with caution)
- Michael Gerlich, “AI Tools in Society: Impacts on Cognitive Offloading and the Future of Critical Thinking”, Societies, 2025 (666 participants, correlation, no causation)