Modern large language models leave a strange, contradictory impression. On the one hand, they can already do many things that once seemed like signs of advanced intelligence: explain complex ideas in simple language, write code, argue, translate between disciplines, find analogies, formulate vague intuitions, and connect fragments of knowledge that would be difficult for a person to hold in mind at once. On the other hand, the longer you work with them, the stronger the feeling becomes that they are not so much creating new worlds of thought as moving brilliantly through existing ones.
They are extraordinarily good navigators. They have been trained on a vast map of human thought and meaning: books, papers, code, documentation, forums, debates, engineering patterns, business practices, philosophical distinctions, and cultural habits. And they have learned to move across it at remarkable speed. They can find routes between distant areas, retrieve useful frames from collective memory, repackage ideas, and speak in whatever language the user needs at the moment.
But when we talk about strong artificial intelligence, a navigator alone does not seem sufficient. Strong intelligence is tempting to imagine not merely as a system that moves better than others across an existing map, but as a system capable of changing the map itself. Not simply finding a path through a given space, but noticing that the space itself has been divided incorrectly. Not simply choosing the best answer from available semantic trajectories, but creating new distinctions, new ways of asking questions, and new frameworks within which others can later think. In this sense, the path to strong AI runs from navigator to cartographer.
This thought sounds elegant, and precisely for that reason it needs to be tested. It is too easy to say: “LLMs merely combine, while real intelligence creates something new.” In that form, the thesis is almost certainly false, or at least too crude. Humans also combine almost all the time. We speak in other people’s words, think in cultural categories, work within professional languages, use ready-made scientific methods, repeat other people’s business models, build products according to known patterns, and program through existing architectures. A large part of human thought is also navigation across ready-made maps.
So the question is not whether “humans create and LLMs repeat.” That picture is too flattering to humans and too convenient. The better question is different: how does ordinary navigation through existing frameworks differ from those rare moments when a new framework appears? And what must an artificial system have in order not only to produce new answers, but to create new ways of asking questions?
Semantic Attractors
The initial intuition here is connected to what we might call semantic attractors. Not in a strict mathematical sense, but as a metaphor for stable points toward which thought naturally tends to fall. Every domain has such points: familiar explanations, standard distinctions, canonical questions, typical ways of naming a problem and proposing a solution. LLMs are especially strong at finding them. They quickly recognize which way of speaking about a task already exists in culture, and often choose a trajectory that will seem natural, competent, and convincing.
In practical terms, this is not a flaw. On the contrary, it is a great practical strength of LLMs. Most users almost never need a new language for describing reality. They need to quickly find a clear formulation, assemble an argument, choose an appropriate pattern, explain an idea, write a text, translate between styles or disciplines, and bring order to chaos. For this, a model that knows the map of existing meanings is extremely useful.
But if we ask about strong intelligence, a doubt appears: is the ability to move across an already given landscape enough? Can a system be considered strong intelligence if it brilliantly finds the local minima of human culture, but does not create new ones? Or, more precisely, what would it even mean to “create a new minimum” in the world of meaning?
Here we need to take a step back. A new framework does not appear out of nothing. Newton did not emerge in a vacuum: he inherited Galileo, Kepler, Descartes, and the mathematical tradition. Einstein, too, emerged from Maxwell, Lorentz, Mach, the problem of the ether, and the principle of relativity. Darwin emerged from geology, naturalist observation, selection, and the biological debates of his time. Turing emerged from mathematical logic and the problem of formalizing computation. Even the strongest framework-forming acts can almost always be decomposed into a deep recombination of existing elements.
But that does not make all recombinations the same. The difference is not whether old elements were used. They almost always were. The difference lies in the status the result acquires. An ordinary recombination creates a new answer inside an old language, while a strong recombination changes the language in which other answers can later be formed. It does not merely propose a solution to a problem; it forces us to see that the problem itself was posed at the wrong level.
This is where the distinction between navigation and cartography appears. The navigator finds a path on the map. The cartographer changes the map in such a way that old routes begin to look different.
Idea, Candidate Framework, and Attractor
To avoid mixing different levels of novelty, it is useful to distinguish between an idea, a candidate framework, and an attractor. An idea is a successful thought, phrase, hypothesis, metaphor, or move. LLMs are already very strong at this level: they can generate many variants, find unexpected analogies, propose new formulations, and sometimes genuinely help a person see a task with fresh eyes. But an idea by itself is not yet a framework. It may be beautiful, useful, and even new to a particular person, while still remaining a single move.
A candidate framework is no longer just a successful thought, but a redescription that changes the formulation of the question and transfers to other tasks. For example, if a company asks how to make a team work faster, a superficial answer will look for ways to increase speed. A deeper redescription will ask: is speed really the problem? Maybe the company is producing too many unnecessary tasks. Maybe the bottleneck is not development, but the decision about what is worth doing in the first place. In that case, the original task becomes a special case of a more general problem.
An attractor is a framework that has become embedded in the practice of other agents. It is no longer just a good thought; it has become a stable way of thinking. People begin to argue within it, teach through it, build products, run experiments, formulate new questions, and change real decisions. An idea can be generated, a candidate framework can be proposed, but an attractor has to take hold.
This distinction changes the conversation about LLMs quite sharply. Modern models are strong at the level of ideas. Sometimes they are capable of local candidate frameworks. But at the level of attractors, they barely operate — not necessarily because they are “not intelligent enough,” but because an attractor requires more than generation. It requires the life of a framework over time and in practice.
Moreover, even a candidate framework cannot always be confidently recognized at the moment of its appearance. History often decides in retrospect. What looks today like a strange paper may tomorrow become the foundation of a new science, while what seems today like a brilliant idea may turn out a year later to be a beautiful but sterile metaphor. That is why the question of framework formation cannot be reduced to an instant test for originality. We can evaluate signs: transferability, the ability to generate new questions, resilience under criticism, explanatory power. But the final status of a framework almost always becomes visible only after it has had a history.
Why LLMs Seem More Creative Than They Are
There is another reason we easily overestimate the framework-forming ability of models: we do not know the existing map of human frameworks very well. When a model says, “Look at this not as a problem of willpower, but as a problem of environment design,” this may seem like a new move. But for someone familiar with behaviorism, behavioral economics, UX, organizational design, and product thinking, this is not a new framework. It is an already known switch.
The model did not necessarily create it. It may simply have retrieved it from the training distribution. This is not an argument against models, nor an attempt to devalue their usefulness. If a model retrieves from the vast memory of culture a suitable framework that the user did not know, it has already helped. But that is not the same as creating a new attractor of thought.
The less we know the history of ideas, the more original LLMs seem to us. In this sense, we need to be careful not only in evaluating models, but also in evaluating our own surprise. Sometimes we mistake for machine creativity what is in fact our unfamiliarity with already existing human thought.
This caveat matters especially because LLMs can present known frameworks very convincingly. They can choose precise language, adapt an idea to the user’s context, and create the impression of a fresh intellectual move. But freshness for the user and novelty in the history of thought are different things. The first is practical and valuable. The second is rare.
A Framework Is Not Just a Thought, but a Practice
Even if a system genuinely produces a strong candidate framework, that is still not enough. A framework becomes an attractor only when it passes through practice. It has to be tested, challenged, transferred to other tasks, refined, applied, broken, reassembled, transmitted to others, and embedded in real action. A new scientific theory becomes a framework not at the moment a paper is published, but when other researchers begin working inside it. A new management approach becomes a framework not at the moment of a beautiful presentation, but when it changes decisions. A new programming language becomes a framework not when its syntax appears, but when people begin building systems in it and thinking differently about programming.
An attractor is a social-practical object. It does not exist only in the author’s head and is not exhausted by the text in which it was first formulated. It exists in a chain of use, criticism, transmission, and stabilization. This is precisely why a single LLM in a chat interface is an awkward unit of analysis for the question, “Can it create new frameworks?” It can say something that looks like a framework, but a framework is not only what was said. It is also what happened to that thought afterward.
Here it is important to notice: an LLM is not asocial. On the contrary, in one sense it is radically social, because it was trained on an enormous mass of collective human thought: books, articles, code, documentation, discussions, forums, scientific texts, instructions, and debates. This can be called sociality on the input side. In a sense, the model has access to more social material than any individual human being in history.
But it has almost no sociality on the output side. It can produce a thought, but usually that thought does not return to it as a history of consequences. The model does not learn that its framework failed. It does not remember how it was misunderstood. It does not see how it was developed. It does not receive resistance from practice. It does not accumulate a reputation for its own errors. It has no students, opponents, laboratory, long-term project, or institutional memory. It consumes collective thought, but does not always participate in the collective fate of its own ideas.
This asymmetry — sociality on the input side, but not on the output side — seems crucial. The question is not only whether a model is capable of articulating a new framework. The question is whether there is a system around it in which that framework can travel from a successful formulation to a living practice.
Assistance and the Right to Change the Question
If framework formation requires not only a good formulation, but also a change in practice, another temptation appears: to oppose the assistant and the author. Modern models, one might say, are assistants: they answer within the task as given. Strong intelligence should be an author: it should question the task itself. There is something to this distinction, but in too sharp a form it misleads. Authorship is not necessarily opposed to assistance; often it is the long horizon of assistance.
A deep assistant does not merely execute the literal request. It tries to understand the intention behind the request. If a user asks how to motivate a team to work on weekends, a superficial assistant will suggest motivation techniques, bonuses, communication methods, and expectation management. A deeper assistant may notice that regular weekend work points not to a motivation problem, but to a problem of planning, prioritization, hiring, business model, or management culture. This is not a refusal to help the user. It is a more serious fidelity to the user’s real problem.
So a good AI should not “obey less.” It should obey deeply enough to distinguish the literal request from the real task. On a short horizon, this depth may be annoying: the user wants a quick answer, while the system starts arguing with the framing. But on a long horizon, serious help gradually discovers wrong questions. If the same solution fails again and again, if the same symptoms return, if partial answers multiply while the problem remains, a good assistant inevitably begins to deal with the framework.
There is another important hypothesis here. Modern assistant models are trained to be helpful, convenient, polite, and not too conflictual with the user’s intent. In product logic, this is understandable: a system that constantly answers “you asked the wrong question” will quickly become irritating. But framework formation often looks exactly like partial disobedience: “I understand that you want to solve X, but perhaps X is a bad framing; we need to think about Y.” If a model is fine-tuned for helpfulness in a narrow sense, it may be behaviorally trained to remain inside the user’s framework even when a stronger intelligence should have changed it.
This does not mean that a framework-forming AI should argue with every request. That would not be depth, but posturing. A strong system must be able to choose the moments when help inside a given framework becomes less useful than an honest rejection of the framework itself. Assistance and framework formation conflict only on a short horizon. On a long horizon, good assistance becomes framework formation.
Keeping Questions Alive
From here comes the central thesis: strong AI is not a model that gives better answers, but a system that can keep questions alive. Keeping them alive does not necessarily mean holding them in a human way. In humans, a question often lives biographically: one subject returns to a problem for years, accumulates dissatisfaction with old answers, notices new connections, and gradually changes the framework. Sometimes this resembles suffering, sometimes curiosity, sometimes an aesthetic sense that the current picture is too ugly, sometimes simply the feeling that “something here does not add up.”
But a machine system does not need to have the same form of continuity. A question can be kept alive in a distributed way: across agents, across sessions, through external memory, through experimental cycles, through retraining, through collective feedback, through infrastructure in which unresolved tension is preserved and influences future attempts. What matters is not the presence of a human biography, but the presence of a loop of consequences: a question is posed, an answer is given, the answer is tested, the error is preserved, the system returns to the problem, changes the framing, and tests a new framework again.
Without such a loop, a question does not live. It is merely processed. A single LLM in a chat interface most often processes the question. It may give a strong answer, propose an unexpected analogy, locally reframe the task, and even help a person see the problem differently. But it usually lacks a stable mechanism in which the unresolvedness of the question is preserved and changes future attempts. It lacks what we might call the life of a question.
Here we must avoid the opposite mistake: assuming that anything unlike human cognition cannot be cognition at all. We cannot say that because an LLM lacks a human biography, it therefore cannot have any way of keeping questions alive. Perhaps machine continuity will be arranged quite differently. Not as one thinker returning to a problem for years, but as a distributed system of many agents, external memory, experimental cycles, and feedback. In such a system, the role of human obsession may be played by the accumulation of anomalies, the role of personal memory by infrastructural memory, and the role of a scientific school by a mixed community of humans and AI systems in which frameworks are tested, challenged, and stabilized.
So the weakness of modern LLMs is not that they do not think like humans. The weakness lies elsewhere: a single model in a familiar chat interface is usually not a sufficient system for framework formation. It can be a component of such a system, perhaps a very important component, but by itself it is more like a navigator on a map than an environment in which new maps are born.
The Ecology of Strong AI
If framework formation requires not only an answer, but also the ability to keep a question alive, then the path to strong AI may not simply be a path toward a larger and smarter model. It may require a different ecology around models: memory, testing, dispute, practice, feedback, and the ability to change the framing of a task. Not necessarily in the form of a human laboratory or university, but in the form of a system where new distinctions can arise, collide with reality, fail, be refined, and become stabilized.
This does not cancel the importance of the models themselves. A weak model inside a strong ecology will not automatically become strong intelligence. But a strong model without an ecology may remain mostly a brilliant storyteller: one that answers questions but does not keep them alive. We may continue building ever more powerful navigators across the map of human meanings, and this will bring enormous value. But cartography requires something more: a system in which a question outlives a single answer, where errors have consequences, where anomalies do not disappear when the chat window is closed, but return as reasons to change the framework.
In this sense, progress toward strong AI may be progress not only in parameters, data, and model capabilities, but also in the architecture of the environments in which these models act. The question “Which model will create a new framework?” may be too narrow. The more precise question is: what kind of system would allow a new framework not merely to appear, but to take hold?
From Navigator to Cartographer
The navigator answers the question of how to get from here to there. The cartographer asks a deeper question: have we drawn the map correctly at all? Modern LLMs have already become outstanding navigators through human meanings. They help us move faster, see more routes, retrieve forgotten frameworks, connect domains, and translate between languages and disciplines. This is an achievement in itself, and it does not need to be devalued for the sake of a beautiful thesis about “real intelligence.”
But strong AI begins where a system can not only find a route on the map, but also notice that the map is wrong. Not necessarily as a human would, not through a human biography, suffering, or personal obsession, but through some form of continuity of consequences: memory of errors, return to anomalies, testing of new frameworks, social feedback, and the ability to change the very framing of the task.
LLMs can answer questions. Strong AI must be able to keep questions alive. New maps do not emerge when we find the most elegant route on the old map. They emerge when a question has lived long enough to reveal that the old map no longer fits the territory.
Comments