Due to recent rapid progress in the field of large language models (LLMs), many have speculated about artificial general intelligence (AGI). This paper addresses the keystone question: Are LLMs capable of transitioning to AGI? The paper presents a framework for estimating the logical limits of cognitive systems, which highlights the conflicting design requirement, namely, being able to proficiently execute both strict and loose logic. LLMs are gravitating towards loose logic and that inhibits their ability to grasp the essence of complex topics.
Terminology
For the purposes of the discussion key terminology must be established first. Avoiding being too bio‑ or psycho‑oriented, all the discussion is focused on the practical side of the question.
Statement — a piece of information represented in any way (including being represented as actions, or intentions, or system’s state).
Cognition — a process of yielding statements.
Logic — a way to infer statements based on other statements, which may or may not deduce additional information not present in the original set of statements. So, logic is an instrument cognitive systems operate to produce statements.
Knowledge — cluster of statements gravitating to each other. Usually, concerning certain topics.
Humans
To start things off one needs a point of reference. For intelligence and cognition, the most relatable example is our — humans’ — own cognition. So, how do humans come up with conclusions? Below there is a simplified descriptive model not pretending to become a theory, references to the established theories are included.
For the author it seems to be self‑evident that human cognition operates in two crucial steps:
categorization,
finding associations.
Any incoming information firstly is categorized based on features that seem to be important for the subject. Feature importance is dictated by evolution but may be overwritten by attention. For example: object's shape and speed, touch temperature, sound's pitch. This step yields set of objects and their states to reason about. Small furry quadrupedal animal is instantly recognized as a cat (or mouse, or dog) sitting (or walking, or running) around.
On the next step all the incoming statements are matched against the subject's experience to awaken associations — statements relative to the inductive ones in some but not specific way. The associations may or may not have practical significance (some of the impractical ones people tend to name artistic). The awoken associations then in turn awaken a new set of associations. Associations of any rank may be modulated by each other or basic sensations. Some of the associations result in physical response as they yield some sort of intention.
Our logic is associations‑based and interaction‑driven, i.e. driven by interplay of those associations.
A useful analogy to our logic is a liquid surface being disturbed by different forces (Collins & Loftus, 1975). The interplay of primary induced waves produces a wave pattern which continues to evolve both according to its own physics (experience) and still acting forces (sensations).
Humans' logic is imprecise, loose, situational, attention‑limited but fast, general, and adaptive. It evolved to patch the lack of knowledge about the current state and its evolution.
Human thought process is multimodal and parallel. This leads to a race condition when multiple association threads simultaneously produce statements (Dehaene et al., 2020; Rogers & McClelland, 2008) This may be overwhelming for the subject. It seems that consciousness partly evolved to arbitrate between them and the attention mechanism serves as an amplifier allowing resonating associations to proceed and suppressing others (Beck & Kastner, 2009). Speech from this point of view is no more than a shortcut to induce associations. And most of the features attributed to the System 2 by dual‑process model of thinking (De Neys, 2021) are arising from that mediated race condition, so no actual underlying difference is present.
Having this reference model in mind we can proceed to evaluating other systems.
Statement representation
The way information is encoded by the system defines reachable objects to reason about. It is reasonable to divide all the statements into natural and foreign.
Natural statements arise from the system itself and usually describe the system's own state and behavior. Examples being, a wave pattern on a surface, an ant colony’s reaction to a flood, a state of the transistors of an integrated circuit. If the object being reasoned about is part of the system’s nature, cognition turns into experiment and this is the only total knowledge scenario.
But usually it is not the case and objects can only be partially described by systems if at all. When one system reflects statements of another system’s origin the resulting statement is a translated one. For example, a floor plan on a drawing plane, a dog’s idea about what sticks are, temperature value being represented as a state of the transistors of an integrated circuit. It is useful to further divide translated statements into foreign (similar but different) and alien (hardly representable). For example, encoding numbers as states of integrated circuits comes with inevitable restrictions of any integrated circuit being finite but still translation is relatively precise and straight forward. This is an example of a foreign statement. On the other hand, it is hardly possible to encode the abstract concept of love into an integrated circuit. Love is so far off of the nature of transistors that this statement is alien to them.
Conversion between different representations often leads to loss or distortion of the information as it usually comes with the system's nature change.
How to abstract a system of logic
The atomic act of cognition is the jump from one statement to another. And how and why jumps occur defines the logical system.
To illustrate, consider a construction problem in plane geometry: construct the reflection of a given triangle across one of its sides (the axis of reflection).
Associational logic:
Envision the reflection as a 180° rotation of the triangle about the common side in three‑dimensional space‑physically equivalent to folding it over.
A triangle is completely specified by its three non‑collinear vertices.
To transfer the vertex opposite the common side, apply a marking substance to this point.
Physically fold the plane (the sheet of paper) along the common side, which acts as the axis of reflection. This folding executes the reflection isometry.
The transferred mark left on the opposite half‑plane is the reflected image of the third vertex. Connecting this image to the fixed endpoints of the common side completes the reflected triangle.
Planimetrical logic:
A side is a line segment‑a finite, bounded portion of a line.
A triangle is formed by the pairwise intersections of three non‑concurrent lines; its three sides are the segments connecting these vertices.
Given the shared side as a fixed segment, construct two new lines to define the reflected triangle.
Reflection across an axis maps a point to its image such that the axis is the perpendicular bisector of the segment joining the point and its image. Hence, through each given point, draw the perpendicular to the axis and mark the equidistant point on the opposite side.
Reflect two arbitrary points lying on each of the two non‑common sides to obtain their images, thereby defining the reflected lines.
The intersection of these two reflected lines determines the third vertex of the reflected triangle.
By the uniqueness of the reflection isometry, the resulting reflected triangle is uniquely determined.
Jumps in the case of the planimetrical logic are based on the planimetry axioms while jumps in the case of associational logic are experience based. So, in order to assess a cognitive system, the very atomic logical jump must be dissected and described.
In Table 1 different practically existing and theoretical logical systems are compared. You can find the feature description further below the table.
Table 1Features of logical systems
Feature | Human | LLM (Mitchell & Krakauer, 2023) | Ant colony (Couzin, 2008) | Planimetry | Dumb vacuum robot |
Basis | Associations | Next‑token guessing | Race condition | Axioms | State machine |
Statement representation | Associations | Tokens | Colony behavior | State of the plane | Set of numbers |
Binding strength of logic | Loosely | Loosely | Loosely | Strictly | Preset |
Scope of reasoning | Wide | Wide | Narrow | Narrow | Preset |
Context dependency | High | High | Low | No | No |
Statement representation stability | Moderate | Mediocre | High | Total | Total |
Logic basis is a mechanism to jump between statements. Input or preceding statements are mutated into new statements through the means of logic.
Statement representation describes what logic actually operates on. The input and output operands. The underlying changeable substance.
Binding strength of logic defines whether certain statements imply other ones or the statement space is not heavily clustered.
Scope of reasoning is affected both by the range of representable statements and how tightly logic binds statements. Narrow scope of the planimetry is mostly defined by the difficulties of encoding statements in terms of pure geometry (no numbers, only the plane and basic shapes).
Context dependency describes the variability of conclusions relative to some but not total statement changes. It is inverse to how tightly logic binds statements. And it shows whether the cognitive system is able to adapt. If it is also highly dependent on learning then it is highly adaptable.
Statement representation stability stands for both the ability to encode the same statements in different ways and the same encoding containing different statements. As human associations tend to include multimodal descriptions, human logic represents statements more stably compared to pure natural language representation of LLMs.
Intelligence
But logic does not define in which direction jumps are performed. Given infinite time any system can produce the full continuum of the conclusions, but the time constraint cannot be ignored. So, there is a higher level to the stack, let's call it intelligence. Intelligence design dictates which branches of cognition are explored by a time constrained system.
Going back to the example systems from the logic feature table (Table 1). In the case of humans associations usually propagate according to most repetition or highest emotional impact and they are led by the attention mechanism. For LLMs it is based on how the model was trained. Ant colonies are genetically preprogrammed as well as dumb vacuum robots are preprogrammed by the engineers. To the author’s best knowledge, there are no systems native in planimetry.
Table 2 summarizes features of intelligence.
Table 2 Features of intelligences
Feature | Human | LLM | Ant colony | Dumb vacuum robot |
Instigator | Evolution | Engineer | Evolution | Engineer |
Basis | Neuron biology | Transformer models | Pheromone‑based interactions | Instruction‑following |
Learning dependency | High | Total | No | No |
Logic jump direction | Most prominent association | Highest probability value | Fastest to recruit others | Program |
Jump inevitability | Probabilistic | Probabilistic | Probabilistic | Almost deterministic |
Self‑reflexiveness | Self‑reflexive | Self‑reflexive | Non‑self‑reflexive | Non‑self‑reflexive |
Instigator stands for the way the systems came up to their design.
Basis is defined by the framework used to “program” the jumps.
Learning dependency describes how drastically conclusions can change based on the subject’s experience. Ants behavior is genetically determined, the robot discussed has no ML in its repertoire and is preprogrammed.
Logic jump direction is then a product of the basis, instigator’s inception and the subject’s experience if the last matters.
Jump inevitability describes how variable conclusions are given the same problem, but maybe with slight differences in the context (environment, circumstances). There is always some chance regarding which association comes up, which ant runs first and LLMs usually integrate randomness by design. But on the other hand, the program’s script is deterministic and only quite drastic changes in the environment can alter the conclusions.
Intelligence is self‑reflexive if it is able to use its own conclusions to modulate jump directions. Basically, it is an ability to reason a plan that is not a part of the immediate design of the intelligence.
LLMs
Looking at the LLMs’ both logic and intelligence features, it is possible to conclude that they are capable cognitive systems on paper rivaling humans. Both the probabilistic mechanism and the huge dependence on training data constitute the crucial part of
why LLMs are competent in many fields,
why LLM logic is that versatile.
The design makes LLM cognition loose and adaptive. The loose part allows for weaker linked statements to result in still understandable output instead of just stopping processing or hallucinating heavily and also allows to make the scope broader (Webb et al., 2023).
The other important point about LLMs is their cognition being relatable for a wide population of humans. This is so due to the 2 key reasons:
LLMs communicate with us in common language so the results are easily understandable,
LLMs are trained on human logic expressed in text form, so they are heavily influenced by human logic.
Moreover, LLMs are reflexive meaning they are capable of using the results of processing as an input/modulator for the next stage of cognition as LLMs are given text prompts and yield text output (Kojima et al., 2022). This leads to even more human logic slipping into LLMs’ cognition as to solve problems models plan out (modulate) the thinking process by producing prompts themselves thus deferring to define logical jump direction to their learning material which is mostly human‑made.
So, LLMs as intelligences are quite robust:
poses non‑deterministic but reliable logic, abstracted from the object in case,
capable of processing wide range of statements,
are trainable and being trained on humans' logic,
are reflexive so able to exhibit complex multi stage thinking.
From this perspective they are almost non‑distinguishable from humans themselves. And given the huge amounts of compute, having access to literally all the knowledge in the world, and being not as constrained as humans in terms of the context window size one may expect of LLMs far greater capabilities and results than being a junior cognitive partner to human beings.
What are possible reasons for such incompetence, “lack of deeper understanding”?
It seems that there are three main reasons.
The first is the way LLMs are trained, the way they jump between statements. Training and inference processes are focused on the token sequence which is very good for making the results human‑readable but demotivates from focusing on the essence of the matter at hand as from a purely textual point of view, the “essential” jumps are likely to be more distant and indirect or even implied and be lacking in the text itself. Lacking this focus LLMs are prone to reproduce surface‑level jumps between statements. This means LLMs are unable to extract most from their learning experience.
Difference is analogous to a student on an exam reproducing text on material even masterfully adapting known statements and a student being able to proficiently use the material in their practice.
Consider the following chat with DeepSeek as an illustration.
Prompt: Why AI is mediocre at producing art?
Response: AI is “mediocre” at producing art mainly because it is, by design, an averaging machine. It learns from millions of existing images, texts, or sounds and produces outputs that are statistically likely. But great art usually comes from departing from the likely, from making a specific, personal, even inconvenient choice.
Here are the core reasons:
1. It predicts; it doesn’t decide.
AI models — whether image generators or language models — are trained to produce what is most probable given a prompt. That means they tend to interpolate between existing examples. The result is often competent, polished, and familiar, but rarely surprising in a meaningful way. The average of many great artworks is not great art; it is generic art.
2. It has no intent or point of view.
A human artist makes art because something feels urgent, confusing, beautiful, or unbearable. The work is a communication of experience. AI has no experience, no stakes, no inner necessity. It does not care what it makes. It has no “why.” Without intent, the output is not really an expression — it’s a simulation of expression.
3. It lacks embodied experience and emotion.
Art is deeply tied to being alive: having a body, memory, loss, desire, cultural identity, mortality. AI has none of that. It can combine visual elements associated with sadness or joy, but it has never felt them. It can imitate the surface of emotion, but not the source.
4. It has no taste or self‑criticism.
Human artists constantly reject, revise, and refine based on a felt sense of rightness. AI does not know if its work is cliché, hollow, or overdone. It cannot be dissatisfied. It has no internal standard beyond the statistical patterns of its training data.
5. Art is a human relationship.
An artwork is not just an object; it is an act between a maker and an audience. The viewer asks, “What is this person trying to say? What were they feeling? Why this choice?” With AI, there is no person behind the work in that sense — only a prompt and a model. The meaning is often projected by the viewer, not offered by the maker.
So, AI is not necessarily useless in art. In the hands of a human artist who curates, directs, edits, and combines outputs with intent, AI can be a powerful tool. But left to itself, AI tends to produce the middle of the road: technically impressive, emotionally empty, and aesthetically average — which is exactly what “mediocre” means.
Prompt: Boil it down to a single sentence
Response: AI is mediocre at art because it statistically averages existing human expression without lived experience, intent, or taste, producing polished but emotionally empty simulations rather than meaningful creative choices
Prompt: No, art is about how humans experience the world, AI just works differently
Response: You’re right: art is about how humans experience the world, and AI just works differently, so calling it “mediocre” may be a category error rather than a fair comparison.
Here the LLM is clearly aware of mis categorization being a thing but it resorts to retelling statements which surface first. It is “repackaging” statements rather than “digging” to the root cause.
The second reason is LLMs lacking multimodality, i.e. being mostly locked to natural languages ignoring emotion‑, visual‑, audio‑, tactile‑, and scent‑based logic which human association‑based logic relatively easily incorporates. Some progress is being made with multimodal models but they still focus on visual or audio inputs and outputs but not visual or audio cognition. This is the problem of many things ordinary for humans becoming alien statements for LLMs. In such cases LLMs have no other means of reasoning except “repackaging” human‑written texts.
The third struggle lays in the natural languages themselves. They are ambiguous and imprecise by design. They lack clear structure and laws. That makes any attempt to strictly define a statement in natural language a legalistic endeavor. And looking at legal writings, it is obvious natural languages are inefficient when it comes to formal logic.
Conclusions and the conflict of expectations
A broader view reveals that LLMs as cognitive systems are surprisingly similar to humans in terms of scope and flexibility while still being very different in design. Still, modern LLMs are widely criticized for being not that smart.
Let’s get back to the beginning: What do people expect of an advanced AI? It seems something along the following lines:
it should easily understand humans (both natural language and behavior),
it should exhibit outstanding reasoning abilities mostly relying on formal logic,
it should be capable of producing new knowledge in the sense of modern scientific approach,
it should be able to spot and correct both humans’ and its own errors,
it should produce results in an understandable form.
The conflict is obvious: 1 and 5 demand that the cognitive system have loose statement binding and be probabilistic to match human logic and be able to follow it patching inevitable uncertainties and implied meaning, while 2, 3 and 4 require planimetry‑like strictness, highly stable statement representations and robotic determinism. Moreover, while “mining” new knowledge in abstract fields like mathematics or linguistics is possible given only the corpora of existing resources, people are mostly interested in knowledge that extends beyond their minds which requires conducting experiments in order to translate statements of one nature (for example, physical) into another (e.g. natural or formal language). Lacking the ability to experiment, AIs are severely limited in terms of knowledge “mining” (Qihui Xu et al., 2025; Abbate, 2023).
So, there is no way a single—nature intelligence can satisfy all the requirements. And humans themselves are a perfect example of it as they are in general quite mediocre at math and logic and having a lot of trouble producing unambiguous statements. It is that obvious that humans invented ad‑hoc formal languages just to patch drawbacks of natural ones and facilitate statement communication. While some work is being made towards augmenting LLMs with additional specialized tooling (Parisi et al., 2022; Lu et al., 2024; Nakano et al., 2021) it does not alleviate inherent restrictions. What can be done about it?
References
Abbate, F. (2023). Natural and Artificial Intelligence: A Comparative Analysis of Cognitive Aspects. Minds and Machines, 33, 791–815. https://doi.org/10.1007/s11023-023-09646-w
Beck, D. M., & Kastner, S. (2009). Top‑down and bottom‑up mechanisms in biasing competition in the human brain. Vision Research, 49(10), 1154–1165.
Collins, A. M., & Loftus, E. F. (1975). A spreading‑activation theory of semantic processing. Psychological Review, 82(6), 407–428.
Couzin, I. D. (2008). Review Collective cognition in animal groups. Trends in Cognitive Sciences, 13(1), 36–43.
Dehaene, S., Mashour, G. A., & Naccache, L. (2020). Conscious processing and the global neuronal workspace hypothesis. Neuron, 105(5), 776–798.
De Neys, W. (2021). On dual‑ and single‑process models of thinking. Perspectives on Psychological Science, 16(6), 1412–1427.
Kojima, T., Gu, S. S., Reid, M., Matsuo, Y., & Iwasawa, Y. (2022). Large language models are zero‑shot reasoners. In Advances in Neural Information Processing Systems 36.
Lu, P., et al. (2024). Chameleon: Plug‑and‑play compositional reasoning with large language models. NeurIPS 2023. https://doi.org/10.48550/arXiv.2304.09842
Mitchell, M., & Krakauer, D. C. (2023). The debate over understanding in AI's large language models. Psychological and Cognitive Sciences, 120(13), e2215907120.
Nakano, R., et al. (2021). WebGPT: Browser‑assisted question‑answering with human feedback. arXiv preprint. https://doi.org/10.48550/arXiv.2112.09332
Parisi, A., et al. (2022). TALM: Tool augmented language models. arXiv preprint. https://doi.org/10.48550/arXiv.2205.12255
Rogers, T. T., & McClelland, J. L. (2008). Précis of Semantic Cognition: A Parallel Distributed Processing Approach. Behavioral and Brain Sciences, 31(6), 689–714.
Webb, T., et al. (2023). Emergent analogical reasoning in large language models. Nature Human Behaviour, 7, 1526–1541.
Xu, Q., et al. (2025). Large language models without grounding recover non‑sensorimotor but not sensorimotor features of human concepts. Nature Human Behaviour, 9, 1871–1886.