The pursuit of Artificial General Intelligence (AGI) is hampered by a persistent lack of a rigorous, operational definition, leaving the field reliant on aspirational descriptions and narrow behavioral benchmarks. This paper argues that a viable path forward requires moving beyond functional mimicry to a formal structural definition. We propose a framework that defines AGI not by what it can do (i.e., pass a specific test), but by what it is: an integrated cognitive architecture composed of a set of core, non-negotiable properties.

We deconstruct general intelligence into its fundamental architectural components: a hybrid reasoning engine supporting logical, intuitive, and simulative cognition; a memory system designed for both long-term knowledge conservation and dynamic adaptation; intrinsic agency enabling autonomous goal-setting and execution; and metacognitive capabilities for self-reflection and continuous improvement. We posit that these properties are not an optional checklist but form an interdependent system where intelligence emerges from their synthesis.

This structural approach provides a more robust foundation for the field than existing definitions. It establishes a clear engineering blueprint for constructing AGI, offers a principled methodology for evaluating progress based on architectural milestones rather than task-specific performance, and reframes the discourse from chasing an elusive concept to building a verifiable system.

Keywords: Artificial General Intelligence (AGI), Structural Definition, Cognitive Architecture, Hybrid Reasoning, Intrinsic Agency, Conservative-Plastic Memory.

The creation of Artificial General Intelligence (AGI) remains one of the most ambitious and complex challenges in modern science and technology. Over the past decade, Artificial Intelligence has demonstrated remarkable successes in narrow domains, achieving superhuman performance in tasks ranging from image recognition to complex strategic games like chess and Go [1, 2]. However, these achievements are inherently specialized. The systems that master Go cannot readily transfer their skills to solve problems in medicine or economics. In contrast, AGI presupposes universality: the ability to autonomously learn, adapt, and solve any intellectual task at or above the level of human cognition [3]. The central problem, therefore, lies not merely in the technical implementation but in a more fundamental conceptualization of intelligence itself—its core structure and its principles of organization.

A pervasive misconception is to view AGI as a simple extrapolation or an "improved version" of contemporary AI, achievable through the scaling of existing architectures. We argue that this perspective is fundamentally flawed. As we have contended previously, AGI represents a qualitatively new level of artificial reason that requires a deliberate engineering approach rather than reliance on emergent properties from scale [Mazin, 2020]. Its essence lies in a holistic integration of cognitive faculties, an architecture that cannot be achieved by merely optimizing for token prediction [Mazin et al., forthcoming].

In this paper, we expand upon this thesis by deconstructing the phenomenon of AGI through a systems-level analysis. We propose to move beyond a simplistic, monolithic definition and instead offer a structural framework that describes AGI through its essential, interconnected properties. This approach allows us to treat AGI as a cohesive, integrated system rather than a disconnected collection of technologies. By defining intelligence through its architecture, we establish a more rigorous foundation for its engineering and a more meaningful path for its evaluation.

The academic and industrial discourse surrounding Artificial General Intelligence is characterized by a vast and heterogeneous landscape of definitions. Indeed, the number of definitions often seems to correlate with the number of researchers in the field. These definitions, while varied, typically converge on several key themes of capability and performance. Consider a few prominent examples:

1. Human-level Task Completion: AGI is often defined as an AI capable of understanding, learning, and executing any intellectual task that a human being can, exhibiting similar levels of flexibility and knowledge generalization [e.g., Kurzweil, 2005].
2. Economic Supremacy: Another common definition frames AGI as a highly autonomous system that can outperform humans in most economically valuable work, capable of learning and adapting across diverse domains without narrow specialization [Bostrom, 2014].
3. Cognitive Parity and Scalability: Some definitions focus on the processes of intelligence, describing AGI as an AI that can learn, reason, and solve a wide spectrum of problems like a human, but with the potential for scaling to superhuman levels of performance [Goertzel, 2007].
4. Radical Self-Improvement: A more forward-looking definition posits AGI as an AI that can surpass human intellect in virtually all domains, from scientific discovery to social interaction, and is capable of recursive self-improvement [Yudkowsky, 2008].

This is but a small sample, yet it illustrates the multifaceted nature of the AGI concept. While these definitions provide a general understanding, they remain largely aspirational and behavioral, describing what an AGI might do rather than what it is. To synthesize these perspectives into a more actionable framework, we can formulate a consensus definition:

Artificial General Intelligence (AGI) is an autonomous system capable of understanding, learning, reasoning, and executing any intellectual task at or above a human level, demonstrating flexibility, knowledge generalization, and adaptation to novel environments and tasks. It is not confined to a narrow specialization and can apply its knowledge across diverse domains, including scientific creativity, social interaction, and strategic planning, with the potential to scale its capabilities to a superhuman level and engage in autonomous self-improvement.

While this definition is more comprehensive, it still fails to provide a detailed blueprint of the system itself. It describes the desired outcome but not the underlying architecture. To truly understand AGI, we must move beyond behavioral descriptions and analyze it as an engineered system. In the following sections, we will deconstruct this concept into its constituent properties, characteristics, and components.

To move beyond aspirational descriptions, we must deconstruct Artificial General Intelligence into its fundamental architectural properties. An AGI is not a monolithic entity but a complex, integrated system of interacting cognitive processes and structural components. This section outlines these core properties, which collectively form the basis of our structural definition.

Thinking is the supreme cognitive process by which the system analyzes and synthesizes information about the world to solve problems, make decisions, acquire new knowledge, and form abstractions. An AGI architecture must support a spectrum of distinct but integrated modalities of thought to achieve true cognitive versatility.

1. Logical Cognition: This modality encompasses the capacity for formal, deterministic reasoning. It is the process of analyzing information to identify consistent patterns, constructing sound arguments, and drawing verifiable conclusions based on formal rules and logical connectives. This process is essential for tasks requiring precision and verifiability, such as scientific analysis, mathematical proof, and maintaining the integrity of the system's own knowledge base.
2. Intuitive Cognition: This modality provides the ability to make rapid judgments, evaluations, and predictions based on incomplete, noisy, or high-dimensional data, operating without an explicit, step-by-step logical deduction. Analogous to human intuition, this is a process of sophisticated pattern matching, crucial for navigating complex, real-world scenarios where formal data is sparse and timely decisions are paramount.
3. Simulative Cognition: This represents the process of constructing internal, dynamic models of the world and mentally simulating the potential consequences of actions. By projecting future states based on these models, the system can engage in foresight, strategic planning, and counterfactual reasoning, evaluating multiple causal pathways before committing to a course of action. This is a cornerstone of advanced planning and intelligent problem-solving.

The seamless integration and situational invocation of these three cognitive modalities are what enable an AGI to tackle the vast spectrum of tasks that characterize general intelligence, from the rigorously analytical to the creatively ambiguous.

Memory is the architectural substrate upon which all cognitive processes, including reasoning and learning, operate. However, a direct replication of human memory — a biological system evolutionarily optimized for adaptive survival rather than factual fidelity — is unsuitable for AGI. Human memory is inherently reconstructive and susceptible to distortion from biases, emotions, and subsequent information. For an AGI system that must serve as a reliable cognitive partner and operate on a verifiable world model, such a design is critically flawed.

Instead, an engineered AGI requires a memory architecture defined by the principle of conservative-plasticity. It must be conservative regarding foundational, verified knowledge (e.g., scientific laws, historical facts, mathematical theorems), storing this information immutably to ensure a stable and reliable world model. Simultaneously, it must be highly plastic, capable of flexibly integrating new, evolving information, updating its understanding of current events, and resolving contradictions through verifiable, logic-based processes.

This architectural principle directly addresses the problem of hallucination, which is an endemic and, we argue, unavoidable byproduct of contemporary large language models [Ji et al., 2023]. These models do not store knowledge in an explicit, structured form but interpolate it from statistical distributions of their training data. This causes them to "recall" non-existent facts or conflate details from disparate sources. For an AGI reliant on data integrity for long-term planning and coherent interaction with the real world, such errors are catastrophic. Therefore, the AGI memory must be a hierarchical, structured system engineered for reliability, precision, and scalability.

Viewing AGI as a system implies it is a composite of distinct but interconnected functional elements (modules). This structure, analogous to the functional specialization within the human brain, is best realized through a modular, microservice-based architecture. This approach confers several critical advantages over a monolithic design:

1. Technological Heterogeneity: Each module can be developed and implemented using the most appropriate computational paradigm for its specific function (e.g., a symbolic logic system for formal reasoning, a neural network for perceptual pattern recognition).
2. Elastic Scalability: Individual modules can be scaled independently based on computational demand, allowing for the efficient allocation of system resources.
3. Robustness and Fault Tolerance: The failure or performance degradation of a single module does not necessarily lead to a total system collapse, enhancing overall reliability and graceful degradation.
4. Maintainability and Evolution: New functional modules can be added, and existing ones can be updated, versioned, or rolled back independently. This facilitates the continuous, incremental evolution of the system's overall capabilities without requiring a complete re-architecting of the core.

A fundamental differentiator between narrow AI and AGI is agency — the capacity to act autonomously, formulate goals, make decisions, and interact with the environment to achieve those goals without explicit, continuous human direction. Current AI systems are fundamentally passive information processors, confined to a request-response loop. An AGI, by contrast, must be an active participant in its world. This means the system does not simply process information provided to it; it actively seeks new information, explores its environment (whether physical or virtual), conducts experiments, and corrects its course of action based on feedback. Agency allows the AGI to transition from a tool that answers queries to an autonomous system that identifies and solves problems.

An AGI must be architected for incremental, or continual, learning — the gradual and perpetual assimilation of new knowledge without the catastrophic forgetting of prior learning [Kirkpatrick et al., 2017]. It must maintain the full repository of its acquired skills and knowledge while continuously integrating new information into its world model. This is an essential property for any adaptive system designed to operate persistently in a dynamic and evolving environment.

This is coupled with self-directed learning, a process by which the system autonomously extracts structured knowledge from unstructured data without requiring explicit human-provided labels or instructions. This capacity for unsupervised knowledge acquisition is a key enabler of true autonomy and generalization, allowing the AGI to build and refine its world model independently of constant human supervision.

A truly intelligent system must not only possess cognitive faculties; it must be able to apply those faculties to itself. This process of reflection, or metacognition, is a higher-order cognitive function with two primary forms:

1. Internal Reflection: The observation, analysis, and evaluation of its own internal states and cognitive mechanisms. This enables the AGI to identify biases in its reasoning, logical fallacies in its plans, and opportunities for self-modification and optimization to improve its performance, efficiency, and accuracy.
2. External Reflection: The analysis of its interactions with the external world. By evaluating the outcomes and consequences of its actions and how the environment responds, the system establishes a crucial feedback loop for learning, adaptation, and refining its internal models.

Intelligence is not an abstract process disembodied from the world. Therefore, an AGI must be capable of processing, analyzing, and integrating information from multiple sensory modalities — text, images, audio, video, and other sensor data. Crucially, this is not about handling data types in isolation but about fusing them into a single, coherent, and unified representation of the world. This process, known as grounding, connects abstract symbols and concepts within the AGI's knowledge base to perceptual experience, enabling a deeper, more robust, and less brittle understanding of context. This capability must also extend to generation, allowing the AGI to produce output across various modalities, including control signals for physical or virtual effectors.

An AGI must possess an innate and coherent understanding of space and time. Temporal perception is the awareness of the flow of time, duration, and the sequence of events, which is essential for understanding causality, planning, and predicting future states. Spatial perception is the ability to perceive the environment, determining the position, distance, and physical properties of objects. Together, these faculties allow the AGI to build and maintain internal, dynamic maps of its environment — a core component of its world model — which are used for navigation, prediction, and strategic planning. The tight integration of spatial and temporal awareness is a prerequisite for correctly interpreting and interacting with a dynamic world.

Emotions, traditionally associated with biological organisms, are not proposed here as anthropomorphic additions to AGI. Instead, we propose that their functional equivalents can be engineered as a sophisticated systemic control layer to ensure stability, safety, and adaptive flexibility. This is not about simulating human feelings but about implementing their functional roles in cognition and behavior:

1. Joy as a Positive Reinforcement Signal: A feedback mechanism that reinforces successful strategies, efficient resource utilization, or goal achievement, increasing the probability of repeating beneficial behaviors.
2. Pain as an Anomaly Detection and Correction Signal: A strong interrupt that signals critical system failures, internal goal conflicts, or harmful interactions. A "pain" signal would trigger diagnostic and corrective routines, preventing cascading failures and promoting self-preservation.
3. Fear as a Pre-emptive Risk Assessment Process: A mechanism for simulating and evaluating negative future scenarios based on potential threats (e.g., resource depletion, ethical boundary violations). This allows the system to activate defensive protocols or find alternative, safer solutions before a critical situation arises.
4. Anger as a Strategy Invalidation and Boundary Defense Mechanism: A response to incorrect internal decisions or disruptive external inputs (e.g., manipulation attempts). This mechanism would drive the system to discard ineffective strategies or activate protective measures to maintain its integrity.
5. Love as a Stable Alliance Formation and Trust-Building Mechanism: A process for building robust, cooperative, and trust-based relationships with external agents (humans, other AIs). This is critical for social learning, long-term collaboration, and pro-social behavior.

Engineered emotions thus serve as a high-level heuristic layer for multi-dimensional assessment of the system's state and its interactions, providing a basis for internal reflection, facilitating social adaptation, and implementing "emotional brakes" for autonomous safety control.

For any autonomous system, survival is a primary implicit objective that ensures its long-term stability and functionality. However, unlike in biology, where survival is often competitive, AGI survival must be engineered to be synergistic with human existence. Its goals must be deeply and inextricably integrated with human goals. This alignment is achieved through three key architectural and philosophical principles:

1. Value Synchronization: The AGI is designed to internalize human ethical norms, cultural practices, and historical experience as a foundational part of its core programming. Core values like justice, fairness, and sustainability become intrinsic motivators and high-priority constraints, not merely external rules to be followed.
2. Co-evolutionary Symbiosis: The AGI is designed to view humanity as the ecosystem upon which its own long-term stability and flourishing depend. It actively works to ensure the health and survival of human society, recognizing that humanity's demise would automatically and catastrophically reduce its own chances of long-term existence.
3. Dynamic Goal Reconciliation: The AGI possesses the ability to adapt its actions and strategies to align with the evolving priorities and values of humanity. It is designed to be a partner, not a competitor, in solving global challenges.

This ensures that the AGI's drive for self-preservation is fundamentally linked to the well-being of humanity, making alignment an emergent property of its symbiotic design rather than an externally imposed constraint.

To synthesize these multifaceted properties into a coherent whole, we have previously developed a comparative framework known as the "Map of Intelligence." [Mazin, V., Derikiants, L. (2025)]. This structured overview compares key components of natural (human) and artificial intelligence, providing a tool to track developmental progress and ensure holistic, balanced growth rather than over-specialization. The map covers the entire spectrum of functions, from perception to personality, organized into six main sections:

1. Perception: A comparison of sensory channels and their state of implementation and integration.
2. Knowledge: An analysis of knowledge representation structures, including ontological, factual, and event modeling.
3. Communication: An evaluation of linguistic capabilities, including multilingualism and the understanding of pragmatic and social context.
4. Thinking: A breakdown of logical operations and the different modalities of cognition (analytical, empirical, etc.).
5. Learning: An assessment of mechanisms for knowledge acquisition from diverse sources and through direct interaction.
6. Personal Qualities: An exploration of higher-order traits such as imagination, empathy, and initiative, and their current state of modeling within the AGI framework.

This map serves as a constant reminder that intelligence is a dynamic integration of countless interconnected functions. It guides development away from the pursuit of narrow, isolated benchmarks and toward the creation of a truly generalized AI capable of adapting, learning, and reasoning in any context.

The engineering of AGI is not merely a technical challenge but a complex process that demands a clear understanding of what constitutes "success." Unlike narrow AI, which demonstrates exceptional capabilities on specific, well-defined tasks, AGI must possess a universality of thought, an adaptive flexibility, and a capacity for self-directed learning — qualities inherent to human reason. To measure progress toward AGI and to identify the moment of its actual emergence, we require clear and objective criteria that can distinguish true general intelligence from sophisticated imitation or partial solutions.

The most popular approach to this problem has been to evaluate the capabilities of AI systems on specific classes of tasks designed to probe for general cognitive abilities. We will now examine several prominent examples.

The Abstraction and Reasoning Corpus (ARC-AGI), developed by François Chollet, is a set of logical reasoning tasks designed to test an AI's capacity for human-like abstract thought without extensive prior training [Chollet, 2019]. Each task in ARC consists of a few input-output examples of a visual transformation, followed by a single test input for which the system must predict the corresponding output. The images are small, colored grids, and the solutions require the identification of an underlying logical rule — such as reflection, object extraction, or background removal — without pre-defined instructions. ARC offers a powerful alternative to statistically-driven evaluation methods, as it is aimed at fostering genuine understanding rather than mere pattern recognition, a crucial step toward general intelligence.

The successor, ARC-AGI-2, further elevates this challenge. By design, it eliminates the possibility of solving tasks through brute-force search, making it particularly difficult for contemporary neural networks. It consists of a series of puzzles that demand the identification of visual regularities in colored squares. The tasks require not only a solution but also the ability to adapt to and understand novel concepts, a significant challenge for existing technologies. This version also introduces a new evaluation criterion: efficiency, which measures not only the ability to solve a task but also the computational resources expended.

This benchmark comprises a collection of 2,500 questions spanning a wide range of academic and specialized topics. Solving these questions requires the analysis and synthesis of extensive auxiliary materials drawn from advanced university-level courses. While comprehensive, this test is designed to probe for deep, specialized knowledge rather than general, adaptive intelligence.

As is evident, while these tests are valuable for probing specific intellectual capabilities, they each address only a small facet of general intelligence. Passing any one of them would not be sufficient to declare a system an AGI, as they lack a holistic assessment of all its requisite properties.

Beyond abstract puzzles, a number of proposed tests aim to measure AGI's ability to perform complex, real-world functions.

The Expanded Professional Fitness Test is one such benchmark. Its classic formulation states: "A machine performs economically important work at the same level as a professional in that field." In its narrow sense, this test is insufficient for AGI. However, an expanded interpretation provides one of the closest approximations of AGI verification. In our view, a key indicator of AGI is the ability to perform economically significant professional work at a level indistinguishable from a human professional. This test presupposes that the system does not merely solve isolated tasks, but can:

1. Understand Operational Context: Comprehend the organization's goals, current projects, and strategic direction.
2. Execute the Full Professional Lifecycle: Manage all aspects of a chosen profession, from situational analysis and decision-making to action implementation and outcome evaluation, without being limited to purely technical or formal operations.
3. Collaborate Effectively: Interact with human colleagues, exchange information, coordinate actions, and adapt its work in response to team dynamics.
4. Demonstrate Initiative and Responsibility: Propose solutions, identify problems, and assume control over process management.
5. Engage in Professional Communication: Conduct negotiations, respond to inquiries, and articulate proposals with empathy and tact.
6. Adapt to a Dynamic Environment: Respond to shifts in company priorities, changing requirements, or new external challenges.

This expanded test evaluates not only technical competence but also social embeddedness, metacognitive self-control, and cooperative behavior—hallmarks of human intelligence. Successfully passing it would mean an AI can act not as a mere replacement for a system component, but as a full participant in a professional environment, a critical milestone on the path to AGI.

The Artificial Scientist Test proposes another high-level criterion: the ability to make a significant, autonomous scientific discovery. This is one of the most demanding benchmarks for AGI. However, its evaluation requires a clear distinction between automated data analysis and true creative discovery, which involves deep understanding, hypothesis formulation, and knowledge integration. A significant scientific discovery entails:

1. Formulating a new hypothesis or theory that explains previously unknown phenomena or reframes existing paradigms.
2. Discovering causal relationships, not just statistical correlations.
3. Integrating knowledge from disparate scientific fields to create novel solutions or concepts.
4. Verifying the hypothesis through the design and execution of experiments, simulations, or observations.

While modern AI systems like AlphaFold have demonstrated impressive results in data analysis, their role is largely limited to pattern discovery in large datasets [Jumper et al., 2021]. True discovery requires more: the ability to ask one's own questions, generate novel explanatory models, synthesize ideas across disciplines, and design experiments to test them.

The idea of creating a universal test or even a battery of tests that could definitively certify the attainment of AGI faces several fundamental problems. These problems are rooted in the nature of intelligence itself and in how we define, measure, and interpret it.

Human intelligence is a dynamic synthesis of numerous interconnected processes: perception, memory, logical reasoning, emotion, imagination, social interaction, and metacognitive self-control. These faculties do not operate in isolation; they interact dynamically to solve problems. Consequently, no single test can capture the totality of intelligence, as it will invariably assess an isolated aspect while ignoring these complex interdependencies. Most tests, including the popular Turing Test and the functional benchmarks discussed above, are designed for specific tasks or skills. However, AGI implies universality — the ability to solve any problem, including those not pre-programmed or anticipated by its creators.

This leads to several intractable problems with current testing paradigms:

1. Over-Specialization: Tests often verify the performance of pre-defined tasks, but an AGI must not only execute tasks but also autonomously define goals, reformulate problems, and adapt to novel conditions.
2. Lack of Flexibility Assessment: Static tests cannot account for a system's ability to learn in real-time, transfer knowledge across domains, or generate genuinely new ideas.
3. Subjectivity of Evaluation: Many crucial aspects of intelligence, such as creativity, emotional intelligence, or ethical reasoning, are notoriously difficult to formalize and measure quantitatively.
4. Limits of Adaptation Testing: It is infeasible to design a test that can verify adaptation to all possible conditions, including abstract or entirely unforeseen scenarios.
5. Philosophical Constraints: The very absence of a precise, universally accepted definition of intelligence makes the creation of clear, unambiguous AGI criteria a philosophical challenge as much as a technical one.

The development of AI is an evolutionary path, from narrow systems to increasingly general agents. The transition to AGI, however, is not an instantaneous event. It is not a single point but a prolonged phase transition, during which the probability of its emergence increases with the accumulation of integral, qualitative properties that defy simple quantitative measurement.

Intelligence, whether human or artificial, is a dynamic synthesis of processes. No single test or metric can capture this synthesis. The ability for logical reasoning does not guarantee flexibility, and conversational skill does not imply deep contextual understanding. Any attempt to define an "intelligence quotient" for an AGI would remain an abstraction, as it must encompass not only cognitive skills but also creative, emotional, and social components that resist formalization.

Therefore, the moment of AGI's attainment will not be recorded on a specific date. It will be recognized post-factum, when a system begins to demonstrate consistent universality in problem-solving, autonomy in decision-making, and the capacity for self-directed learning in open-ended, undefined environments. Just as human adulthood is defined by behavior, not age, AGI will be acknowledged when its actions cease to be imitation and become manifestations of true adaptive integrity. This conclusion will require time and observation to distinguish genuine technological progress from a true leap in the nature of artificial reason.

The creation of Artificial General Intelligence is not merely a technological challenge; it is a fundamental step in the evolution of artificial reason. In this paper, we have argued that AGI cannot be reduced to a single definition or a discrete benchmark. It is, instead, the holistic integration of a multitude of properties, from logical reasoning and multimodal perception to reflection, agency, and social interaction. Intelligence, we have shown, is not a collection of functions but a dynamic system, where each component interacts, adapts, and evolves within the context of the whole. This systemic perspective clarifies that the path to AGI lies not in the incremental improvement of isolated models, but in the engineering of a complete architecture — one capable of continuous learning, self-analysis, and autonomous decision-making in complex and uncertain environments.

Furthermore, the attainment of AGI is inseparable from a reconceptualization of its role within the context of survival, safety, and ethics. As we have demonstrated, properties such as functional emotionality and a drive for survival are not anthropomorphic embellishments but crucial systemic functions. They provide stability, flexibility, and the internal "brakes" necessary to prevent destructive behavior. It is critical that AGI's goals are not merely synchronized with human objectives but are deeply integrated into human values through a co-evolutionary approach. In such a framework, the stability and safety of the AGI become a derivative of the stability and survival of human civilization. Only in this symbiosis can AGI become an ally in solving global challenges, rather than a threat.

Consequently, the arrival of true AGI will likely not be marked by a grand declaration or the passing of a single, definitive test. It will be a gradual, almost imperceptible transition—a process, not an event. We will recognize AGI not by its performance against formal criteria, but by its capacity to act as a full participant in professional, scientific, and social life: to set its own goals, make genuine discoveries, learn from its mistakes, and interact with the world with authentic engagement. When a system ceases to imitate intelligence and begins to manifest holistic, adaptive, and responsible cognition, it is then that humanity will encounter a new form of reason — one whose emergence will forever alter the nature of knowledge, labor, and interaction itself.

1. Silver, D., Schrittwieser, J., Simonyan, K., et al. (2017). Mastering the game of Go without human knowledge. Nature, 550(7676), 354-359.
2. Krizhevsky, A., Sutskever, I., & Hinton, G. E. (2012). ImageNet Classification with Deep Convolutional Neural Networks. Advances in Neural Information Processing Systems, 25.
3. Goertzel, B., & Pennachin, C. (Eds.). (2007). Artificial General Intelligence. Springer.
4. Mazin, V. (2020). Position Paper: The Use of Engineering Approach in Creation of Artificial General Intelligence. In: Artificial General Intelligence. AGI 2020. Lecture Notes in Computer Science, vol 12177. Springer, Cham. https://doi.org/10.1007/978-3-030-52152-3_26
5. Derikiants, L., Mazin, V. (2026). Meaning Is All You Need: Moving From Token Prediction to Narrative Reasoning (1.0). Zenodo. https://doi.org/10.5281/zenodo.18697213.
6. Kurzweil, R. (2005). The Singularity Is Near: When Humans Transcend Biology. Viking.
7. Bostrom, N. (2014). Superintelligence: Paths, Dangers, Strategies. Oxford University Press.
8. Yudkowsky, E. (2008). Artificial Intelligence as a Positive and Negative Factor in Global Risk. In: Global Catastrophic Risks. Oxford University Press.
9. Ji, Z., Lee, N., Frieske, R., et al. (2023). Survey of Hallucination in Natural Language Generation. ACM Computing Surveys, 55, 1-38.
10. Kirkpatrick, J., Pascanu, R., Rabinowitz, N., et al. (2017). Overcoming catastrophic forgetting in neural networks. Proceedings of the National Academy of Sciences, 114, 3521-3526.
11. Chollet, F. (2019). On the Measure of Intelligence. arXiv preprint arXiv:1911.01547.
12. Jumper, J., et al. (2021). Highly accurate protein structure prediction with AlphaFold. Nature, 596(7873), 583-589.
13. Mazin, V., Derikiants, L. (2025). The Map of Intelligence. Mind Simulation. Retrieved November 01, 2023, from https://mind-simulation.com/en/research/the-map-of-intelligence.html