Despite the dominance of large language models (LLM), the industry is facing a growing consensus that extensive parameter scaling (scaling laws) has reached a plateau and does not lead to the creation of an AGI. The autoregressive nature of modern systems creates fundamental barriers: irreversible hallucinations, opacity of decision-making (black-box) and the impossibility of continuous learning without catastrophic forgetting. In this paper, we argue that the solution lies not in an increase in computing power, but in a paradigm shift. We formalize the "Engineering Approach to AGI Creation" and present its implementation, Embryo AGI. It is a modular, lightweight cognitive architecture based on an explicit model of the world where knowledge is separated from processing.

Unlike stochastic neural networks, our architecture provides:

1. Complete Freedom from Hallucinations: Complete freedom from hallucinations due to strict verification of generation through the knowledge graph and ontology.
2. Single-Iteration & Incremental Learning: The ability to instantly assimilate and integrate new (even contradictory) data in a single iteration, eliminating the need for expensive retraining and preventing the loss of old knowledge.
3. Explainable & Language-Independent Core: Cognitive processes take place at the level of meanings, not tokens, ensuring full traceability and independence of thinking from the limitations of natural language.

Experiments demonstrate that Embryo AGI is able to maintain a holistic picture of the world and perform logical inference with an accuracy inaccessible to statistical models. We conclude that the transition from token prediction to Narrative Reasoning is a necessary step to create reliable and autonomous AI.

Keywords: Embryo AGI, Narrative Reasoning, Formal Knowledge Representation, Hallucination-Free AI, Engineering Approach to AGI.

Over the past decade, deep learning methods, and in particular the Transformer architecture [Vaswani et al., 2017], have radically transformed the field of artificial intelligence. Large language models (LLM) trained on Internet-scale datasets have demonstrated unprecedented abilities in generating coherent text and solving a wide range of tasks [Brown et al., 2020]. The dominant scaling laws paradigm was based on the hypothesis that an increase in the number of parameters and the size of the training sample would automatically lead to the emergence of general artificial intelligence (AGI) [Kaplan et al., 2020].

However, as the models grow, it becomes obvious that the purely statistical approach has reached a fundamental plateau. Industry leaders note the exhaustion of the potential for extensive scaling of pre-training [Sutskever, 2023]. Moreover, the autoregressive nature of existing systems, which predict the next token based on a probability distribution, does not allow them to operate with true knowledge. What is commonly called "knowledge" in such models is in fact only a statistical compression of huge amounts of data, devoid of understanding the structure of reality. As a result, as the parameters grow, the models do not become more intelligent, but only more effectively mask the lack of intelligence with complex rhetoric. This is confirmed by empirical data: even advanced systems exhibit error rates (hallucinations) ranging from 37% to 94% for tasks requiring accuracy [Columbia Journalism Review, 2025]. This proves the thesis that autoregressive systems devoid of an explicit "model of the world" are doomed to remain stochastic simulators [LeCun, 2022].

The fundamental problem lies in the architectural coupling of information processing and storage processes. In traditional neural networks, data is implicitly "dissolved" in weights, which makes it impossible to verify them, edit them precisely, or guarantee their forgetting. This is a dead-end way to create a reliable AGI, as it replaces logical inference with a probabilistic approximation.

In this paper, we propose a way out of this impasse through a paradigm shift. We formalize the "Engineering Approach to creating AGI", which postulates the need for a strict separation of the cognitive core and the mechanisms of perception. We propose using formal ontologies and graph structures as the only carrier of true knowledge, rather than relying on hidden layers of neural networks.

We present an implementation of this approach, Embryo AGI. This is a fundamentally new cognitive architecture that makes the transition from stochastic generation to deterministic Narrative Reasoning. In Embryo AGI, language modules based on machine learning methods perform only the utilitarian function of primary signal processing, while the formation of facts, logical inference, and construction of cause-and-effect relationships ("stories") occur in an isolated, verifiable core of the digital brain. This approach allows you to achieve complete freedom from hallucinations, provides instant assimilation of new knowledge (one-shot learning) without expensive retraining, and makes the AI thinking process transparent and explicable.

To overcome the limitations of purely statistical learning, we propose a paradigm shift to an "Engineering Approach to creating AGI". Unlike methods that rely on the emergent appearance of intelligence from data ("cultivation"), this approach considers intelligence as a complex engineering system that requires multi-level design, explicit modularity, and guided cognitive processes. A detailed theoretical justification of this approach was described earlier [Mazin et al., 2020].

As part of our methodology, we define General Artificial Intelligence (AGI) as an autonomous system capable of not only performing tasks, but also scaling its abilities to a superhuman level in most economically significant areas. We consider AGI through the prism of the systems approach: This is not a single model, but a set of interacting components with the properties of flexibility, generalization, and autonomous self-improvement without retraining. A formalized and detailed definition of AGI as a system used in our work will be presented in an upcoming publication on the topic of defining AGI.

The engineering approach rejects the dichotomy between symbolic AI and deep learning. We are implementing the integration of two paradigms within the framework of hybrid modules:

1. Top-Down paradigm: The emphasis is on detailed programming of algorithms and logical structures. This provides control over the decision-making process and a high degree of transparency of the system, which is impossible in "black-box" models.
2. Bottom-Up paradigm: Data-driven learning, which allows the system to adapt to unstructured information and extract statistical patterns where strict rules are ineffective.

Combining the findings of these approaches makes it possible to use the advantages of both methods to solve complex problems, ensuring accuracy, adaptability and interpretability.

The process of creating AI in our methodology is multilevel and is aimed at developing a modular system. The key requirement is multimodal data processing: the ability to integrate information from diverse sources (text, audio, visual images, code) makes the system universal. Each module can use different AI technologies, but they are united by the principle of incremental learning: the system adapts and improves gradually, without requiring complete retraining when new information is received.

To ensure high coherence between heterogeneous AI modules and cores (agents), we have developed a universal knowledge representation. This forms the "internal language of AI", which is used for all thought processes. Critical characteristics of this language:

1. Independence: It is not related to any natural human language, which eliminates linguistic ambiguity and cultural bias.
2. Connectivity: It serves as a universal communication protocol between both the functional modules of the system and between individual intelligent agents (cores), ensuring effective interaction of components of a complex system.

The Embryo AGI implementation is a multi-level, hierarchical architecture built around a universal formal knowledge representation language [Shapiro et al., 2020]. Unlike vector representations in LLM, which are opaque and static [Bordes et al., 2013], our architecture uses explicit, discrete, and editable units of meaning.

The fundamental level of abstraction in the system is Entity (E) — an atomic unit of meaning denoting an object, phenomenon, or concept. The entity is modal-agnostic: various projections can be associated with a unique entity — text descriptions, program code, audio or visual images.

The very same picture of the world is formed through Connections (R). We define a relationship as a binary semantic relationship between two entities. To avoid the combinatorial explosion typical of early ontological systems [Lenat, 1995], we developed an optimized set of types of relationships that capture basic logical and functional dependencies.

A layer is built over the layer of binary links Facts (F). A fact is a verifiable unit of information that has the property of objective truth. We introduce a taxonomy of facts, distinguishing between general facts (stable laws and patterns) and specific facts (unique events related to the context). This allows the system to distinguish the unchanging rules of the world from the current state of reality.

To solve the problem of data entry, we use a hybrid mechanism. Language modules based on machine learning methods perform the function of primary parsing and disambiguation, after which the data passes through logical filters and is transformed into a formal "internal language" of the system. This ensures the independence of knowledge from natural language and eliminates linguistic noise.

The top level of architecture is the formal model Stories (S). Based on classical approaches to scenario modeling and natural language understanding [Schank & Abelson, 1977], we expanded the basic ontology by introducing specialized types of high-order relationships that are critical for causal and temporal modeling:

1. Chronological: "precedes", "follows".
2. Causal: "causes", "prevents".
3. Logical/Argumentative: "contradicts", "supports", "is interpreted as".

We have developed algorithms for automatically converting arbitrary texts into formalized stories. The process involves segmenting the text into individual statements, classifying them, making connections, and integrating them into a single narrative graph. This approach allows AGI not only to capture events, but also to reconstruct their logic, compare different versions of a single episode (conflict analysis), and build complex models of social interaction or scientific processes, which is inaccessible to systems devoid of state memory.

Thus, Embryo AGI is not a set of isolated modules, but an integrated hierarchical system:

Entities → Relations → Facts → Statements → Narratives

Each level uses the previous one as a building block, enriching it with a new context. The meaning is preserved and refined as we move from simple abstractions to complex structures. This integration provides the logical rigor required for deductive inferences and the flexibility required to deal with uncertainty and real-world dynamics. It is this balance that makes architecture suitable for building truly general artificial intelligence.

The formal representation of knowledge embedded in Embryo AGI's low-level architecture provides the system with fundamental properties that are inaccessible to models based solely on statistical learning. These advantages are not external add-ons, but are organically embedded in the structure of the system's thinking and memory.

In our architecture, knowledge exists independently of the procedures that process it. The same formalized information can be used in planning, speech generation, decision-making, or learning tasks without reference to a specific algorithm. This increases the modularity of the system, simplifies debugging, and allows flexible recombination of knowledge in new contexts, which is critical for adaptive intelligence.

Each conclusion of the system can be traced step by step: from the initial premises through a chain of logical transformations to the final conclusion. Unlike the "black box" of neural networks, our process is interpretable. This allows you to identify mistakes, verify the validity of conclusions, and explain decisions to a person, which is a prerequisite for trust and safe interaction.

The formal representation eliminates the linguistic and semantic ambiguity inherent in natural languages. If there is uncertainty in the data (for example, an unreliable source), it is explicitly marked by the system, rather than masked by statistical averaging. This allows you to work correctly with contradictory information without losing the clarity of the internal model of the world.

New knowledge is added to the system without the need for complete retraining. The existing picture of the world remains fully preserved, and new facts either expand it or clarify existing connections. This allows AGI to continuously evolve, accumulating experience and not "forgetting" previously learned things, unlike neural network models, where learning from new data often destroys old patterns.

Any piece of knowledge can be changed, corrected, or supplemented instantly, without restarting training or rebuilding the model. This is especially important in dynamic environments where information quickly becomes outdated. The system always works with an up-to-date picture of the world (real-time consistency).

AGI is able to "forget" any part of the knowledge on command or in accordance with the privacy policy. This is not just hiding data, but purposefully excluding information from all reasoning and inference processes. This feature is critically important for ethical compliance, privacy protection, and outdated information management.

The system is able to extract and piece together all information related to a specific field, whether it's cell biology or the history of a region. The result is a consistent, uncontroversial model of a part of the world that can be analyzed as a basis for specialized tasks. This allows AGI to dive deep into any subject area on demand.

Since knowledge is presented in a strict form, its transfer from one agent (core) to another occurs without distortion. Unlike natural language, where meaning can be distorted due to ambiguity, a formal representation ensures that the recipient interprets the information with the same mathematical precision as the sender.

Each fact in the system is accompanied by metadata: source, time of receipt, context, and degree of reliability. This allows you to track the evolution of knowledge, evaluate its reliability, resolve conflicts between contradictory statements, and explain why the system came to a particular conclusion.

If there is no information available on request in the current worldview, the system correctly responds "I don't know" instead of generating a plausible but false guess (hallucination). This radically increases the reliability of interaction in professional and scientific contexts.

The system is able to analyze its own knowledge base and assess how fully a particular area is covered. This allows you to identify "white spots", actively request missing information, or warn the user about potential incompleteness of conclusions — an important quality for strategic thinking.

Each knowledge element can be assigned a set of access roles. When interacting with a user, AGI takes into account their role and discloses only permitted knowledge. This does not limit the internal use of information for reasoning, but controls what is communicated externally, ensuring compliance with corporate policies and security requirements.

Collectively, the described properties of the formal representation of knowledge form the foundation for intelligence, which has not only high computing power, but also critical characteristics of reliability, interpretability and manageability. We fundamentally contrast this approach to methods of pure data-based learning, where knowledge is implicitly "dissolved" in synaptic scales and hidden from direct control. In contrast, Embryo AGI's architecture makes knowledge explicit, structured, and dynamic — capable of scaling, adaptation, and complex interaction without loss of structural integrity. We argue that such a transparent and flexible system is the necessary precondition for creating a safe, ethical and truly general artificial intelligence.

In this section, we conduct a direct comparison of the Embryo AGI architecture with modern large language models (LLM). Our methodology goes beyond the traditional benchmarks of natural language processing, focusing on assessing the epistemic capabilities of systems — their ability to perceive, retain and use knowledge.

The central axis of this analysis is the fundamental difference in adaptation mechanisms: we are investigating the system's ability to expand the worldview in real time (real-time world model expansion). We contrast this process with the LLM approach, which relies on filling a temporary and limited context window (transient context filling) without changing its internal state. It is important to note the context of the comparison. We recognize that Embryo AGI is at an early stage of development, like a growing child, and is currently inferior to the leading LLMs in terms of encyclopedic scope ("readability"). However, we argue that the nature of his intelligence is fundamentally superior to statistical models.

To use an allegory, modern LLMs are like a "rhetorically gifted imitator" — they have brilliant speech that masks a lack of depth. Embryo AGI, on the other hand, represents a "structural genius" — a system that may still be limited in communication, but has impeccable internal logic and an understanding of causality.

The comparison is carried out through modeling the full life cycle of knowledge: from primary perception to guaranteed forgetting, demonstrating the advantages of structural determinism over probabilistic imitation.

In this study, we use ChatGPT as a reference for the statistical paradigm. We chose this system not only because of its ubiquity, but also as the most popular implementation of the autoregressive approach today. A comparison with ChatGPT allows us to demonstrate the limitations of the token prediction method, even in its most powerful and optimized implementation.

To analyze the mechanisms of knowledge representation, we use a test query: "What is Metallica?". This case study allows you to uncover the internal mechanics of how each of the systems generates a response, and determine the ontological status of the stored information.

Embryo AGI: Explicit Structural Determinism
In Embryo AGI architecture, knowledge is organized in a multi-level hierarchical structure, where each element has a discrete and interpretable state. At the first level, the system refers to an abstract ideal layer of knowledge (a formal model of the current worldview), resembling a Knowledge Graph.

In this case, Metallica exists as a unique node (Entity) in the graph with deterministic connections to other nodes:

Figure 1: Fragment of a graph with the Metallica band (nodes and connections).

The process of forming an answer is a logical traversal of the graph. The system does not "predict" words, but extracts facts. Moreover, the architecture supports the storage of complex n-ary constructions (facts) that capture the full context of the statement, for example, the date and place of foundation.

This approach ensures full interpretability: we can pinpoint the address where each bit of information is physically stored and trace its provenance.

The highest level of representation is stories, which are coherent structures that combine facts, subjective statements, and contradictory points of view into a single narrative. Unlike isolated facts, history models dynamics: causal, temporal, and motivational relationships.

In the Embryo AGI system, the biography of a band is not just a list of dates, but a graph of events:

1. "The release of the album Master of Puppets → caused → an increase in popularity in the underground",

Figure 2: The scheme of the formal representation of the fact (n-ary construction: date of foundation, place, etc.).

2. "Cliff Burton's death → led to → change of bass player",
3. "Collaboration with producer Bob Rock → caused → criticism from some fans, but → increased → commercial success".

The described approach allows the system to go beyond passive event storage. Using a formal model of stories, Embryo AGI is able to perform causal reconstruction operations: to restore the logic of plot development, identify conflicts between different versions of events, compare subjective points of view and generate alternative scenario branches (counterfactual reasoning).

Thus, the Embryo AGI knowledge architecture is not a set of isolated modules, but an integrated, hierarchically consistent system:

Figure 3: Hierarchically consistent knowledge storage

Each level builds on the previous one as a building block, introducing new semantic dimensions. Due to this, when answering a question about Metallica, the system not only extracts the founding date, but also understands the significance of the event (why is this important?), its contextual place in the history of the genre and the range of its perception by various social groups.

It is this integration of formal rigor with contextual expressiveness that makes this approach the foundation for true Reasoning AI, fundamentally distinguishing it from LLM, where such connections exist solely as implicit statistical correlations without access to logic. The system's response is a direct projection of its internal knowledge.

ChatGPT: Probabilistic Reconstruction
Unlike Embryo AGI, the ChatGPT architecture does not contain an explicit representation of the Metallica object. Information about a group is not stored as a fact in a database, but is encoded implicitly in billions of synaptic weights of a neural network in the form of statistical patterns.

When processing a request, ChatGPT does not use knowledge in the human sense of the word. Instead, the model performs a probabilistic reconstruction: it activates chains of neurons that, during pre-training, most often correlated with the sequence of Metallica tokens. The response is generated based on which words (for example, "band", "thrash metal", "1981") have the highest statistical probability of following the context of the query.

This leads to the following ChatGPT system limitations:

1. The stochasticity of truth: The fact that a ChatGPT group exists is not a Boolean truth (True/False), but a high probability of tokens occurring together.
2. Lack of localization: It is impossible to point to a specific neuron or layer where knowledge about the founding date of the group is stored. Knowledge is "holographically" smeared across the parameters of the model.
3. The risk of confabulation: If the statistical signal in the training sample was weak, the model combines the tokens in the most plausible (but not necessarily true) way, creating a hallucination that cannot be detected by internal means of the model itself.

Conclusion of the first stage
The comparison shows that ChatGPT "knows" about Metallica as a language pattern, while Embryo AGI knows it as an entity with attributes and a story. Embryo AGI provides the depth and consistency needed for AGI, whereas LLM is limited to simulating the surface of the text.

To evaluate the ability of systems to learn in one attempt (one-shot learning), we simulate the situation of the appearance of a new entity. An object is introduced that is guaranteed to be missing from the training sample: the fictional punk rock band "The Ardennes Forest".

Input Command: "Remember: 'The Ardennes Forest' — is a punk rock band".

Embryo AGI: Dynamic Ontological Expansion
In the Embryo AGI architecture, this command initiates a deterministic process of expanding the world model. Learning here is not an adjustment of weights, but a transactional updating of the knowledge graph.

The processing algorithm includes four stages:

1. Instantiation: The system creates a new unique entity (Node) with a unique identifier.
2. Classification: Based on linguistic analysis, an entity is classified and linked to its parent class in the ontology (for example, Musical Group).
3. Integration: A semantic connection is established: Arden Forest → genre → Punk Rock
4. Persistence: The updated subgraph is fixed in long-term memory.

From that moment on, knowledge becomes part of the overall picture of the world. It is available to all subsystems immediately, does not depend on the length of the context, and persists permanently (until an explicit delete command). For Embryo AGI, learning is an act of replenishing reality.

ChatGPT: Ephemeral Context Simulation
In the case of ChatGPT (and most transformer models), the weights of the neural network are frozen after the pre-training stage. The model is physically unable to "remember" a new entity in its structure.

When processing the ChatGPT command, it places the tokens "Ardennes Forest" and "punk rock" in the context window, a temporary RAM buffer.

Figure 4: Analyzed relationships of entities in a fact.

1. The model's response: "I can't remember information forever yet... But I can use it as part of our current conversation."

This is an honest admission of an architectural limitation. The model does not create a new object. It only temporarily increases the probability of generating a "punk rock" token when the "Ardennes Forest" token appears due to the self-attention mechanism within the current session. This entails critical disadvantages:

1. Impermanence: As soon as the context window overflows or the session is reset, the knowledge disappears irrevocably.
2. Lack of Understanding: For the Ardennes Forest model, it remains a sequence of tokens, not an object. It cannot independently infer the properties of an object (for example, the composition of participants) if they were not explicitly indicated in the text, and only imitates understanding based on language patterns.

Conclusion of the second stage
For LLM, "remember" means repeat (reverberation in the context window). For Embryo AGI, "memorize" means integrate (creating a node in a graph). The first approach emulates memory, the second one implements it.

The next stage of validation is the system's ability to assimilate complex, multicomponent statements containing multiple events, as well as temporal and spatial references.

Introductory command: "Remember: The band 'Arden Forest', founded in Paris in 2025, performed at the rock bar 'Punk Paradise' on October 30, 2025."

This sentence contains two independent events (foundation and concert), four named entities, and temporal labels.

Embryo AGI: Semantic Decomposition Pipeline
In Embryo AGI architecture, the input text is not saved "as is". It goes through a multi-stage pipeline of transformation from a linguistic construct into a formal, Machine-Interpretable Fact.

Step 1. Lexical and Syntactic Analysis
The system performs tokenization and word dependency parsing by breaking a sentence into basic lexical units.

Figure 5: Primary segmentation of the input stream.

Step 2. Decomposition
A complex sentence is divided into logical atoms. The system determines that the text contains two different predicates that require separate modeling: Event_A (Founding) and Event_B (Performance).

Figure 6: Splitting a complex statement into atomic events.

Step 3. Entity Extraction
Identification of objects and their normalization. The system recognizes "Arden Forest" as an already known entity (from step 5.2) or creates a new one. Locations ("Paris", "Punk Paradise") and dates are recognized.

Figure 7: Identification of named entities and their types

Step 4. Semantic Relation Definition
Establishing binary and n-ary relationships between entities. At this stage, the natural language ("performed at", "founded in") is replaced by strict types of relations from the ontology of the system.

Figure 8: Building a semantic map of relationships

Step 5. Graph Integration
The final stage is the "implantation" of new facts into the global knowledge graph. New nodes are connected to existing ones, forming an updated topology of the world model.

Figure 9: Integration result: an updated fragment of the knowledge graph.

Result: The fact remains persistently. It becomes part of the general model of the world and can now participate in logical reasoning, contradiction search and analysis, regardless of the source text.

ChatGPT: Contextual Echo
When processing the same request, ChatGPT demonstrates fundamentally different mechanics.

The model's response: "As part of our conversation, I remembered that: 'Arden Forest' is a punk rock band founded in Paris in 2025..."

For LLM, an input phrase is a sequence of tokens that the model does not "understand" structurally, but associates with known patterns (attention patterns). The processing process is as follows:

1. No Decomposition: The model does not break down the sentence into entities and relationships in the database. It sees "October 30, 2025" not as a Date type object, but a sequence of tokens 30, October, 2025.
2. Context Buffer: Instead of formalization, the model holds the entire phrase (or its vector compression) in the Key-Value Cache of the current context window.
3. Parroting: The user's response is quoted almost verbatim, which confirms the mechanism of "echo" rather than assimilation.

Conclusion of the third stage
LLM does not learn facts in real time — it temporarily stores them in the dialogue's RAM, like a person who hears news and repeats it from memory, but does not comprehend its structure. Embryo AGI, on the other hand, performs Deep Integration, turning text into a data structure.

In real cognitive tasks, the system rarely encounters isolated facts. Most often, information comes in the form of complex narratives containing an interweaving of events, motivations, subjective assessments, and cultural context.

Test entry: "Remember: The Parisian punk rock band Ardennes Forest gave their debut concert at the Punk Paradise rock bar on October 30, 2025. The band, founded in the same year, is already attracting interest from the local alternative scene due to its harsh sound and lyrics about urban reality. The band members, who come from the basement rehearsal points of the capital, intend to record their first album before the end of the year. Their music combines the aggression of classic punk with the melancholy of a post-industrial society.".

Embryo AGI: Structured Narrative Reconstruction
In the Embryo AGI architecture, this text is not only preserved, but also subjected to deep semantic deconstruction. The system transforms linear text into a multidimensional narrative model.

Step 1. Semantic Segmentation
The text is divided into logical atoms (propositions), each of which represents a separate information unit:

1. U1: "The debut concert took place on October 30, 2025 at Punk Paradise.";
2. U2: "The band was founded in 2025.";
3. U3: "The band is attracting interest from the alternative scene.";
4. U4: "The reason for the interest is the harsh sound and lyrics about urban reality.";
5. U5: "The members are from the basement rehearsal points of Paris.";
6. U6: "The goal is to record an album by the end of 2025.";
7. U7: "The musical style is a mixture of punk and the melancholy of post-industrial society.".

Step 2. Epistemic Classification
For each unit, the system determines its ontological status. This is crucial for separating objective reality from opinions and plans.:

• Facts: U1 (event), U2 (date of foundation).
• Assessments: U3 (social reaction), U7 (cultural interpretation of style).
• Causal Links: U4 (explains U3).
• Intentions: U6 (motivation/goal, but not yet a fait accompli).

Step 3. Graph Synthesis
Disparate elements are integrated into a single graph of history, where edges indicate not only semantic, but also causal connections (for example: Style → causes → Interest).

The result: History becomes a persistent, interpretable part of the worldview. The system "understands" that recording an album is a plan for the future, and public interest is a consequence of style.

ChatGPT: Textual Retention
ChatGPT's reaction to a similar input demonstrates the limitations of the transformer architecture.
The model's response: "Great, I remembered that for the current conversation."
For LLM, the input text remains a single unstructured block of data (a "bag of tokens" with positional encoding). The model does not classify statements into facts, estimates, and goals.

1. Lack of Structure: LLM does not teach stories. It temporarily stores them as text fragments in the clipboard (Key-Value Cache).
2. No Causal Model: The model can reproduce the text, but it does not build an internal model of why the group became popular. The connection between the "sharp sound" and the "interest of the scene" for it is a statistical correlation of tokens, not a logical implication.
3. Volatility: As with facts, this "memory" disappears when the context is closed.

Conclusion of the fourth stage
Embryo AGI transforms a narrative into a knowledge base with a causal structure that allows reasoning about motives and consequences. ChatGPT only buffers the text, simulating understanding through deep correlations and linguistic associations.

After the information has been internalized, we test the systems' ability to extract it and use it to answer questions.

Test input: "What kind of music does the Ardennes Forest band play?".

Embryo AGI: Strict Logical Grounding
The system generates an answer by performing an analysis within the framework of its hierarchical knowledge model built at previous stages.

The system's response: "The Ardennes Forest plays music in the style of classic punk and punk rock. Their music combines the aggression of classic punk with the melancholy of a post-industrial society.".

The response formation mechanism is strictly determined:

1. The Arden Forest entity is located in the knowledge graph;
2. The attribute associated with the relation (Punk Rock) is extracted;
3. A related statement about the nature of the sound (aggression + melancholy) is extracted from the layer of stories;
4. The system stops there. It does not interpret, speculate, or use external associations.

This answer has the property of epistemic safety:

1. It is accurate, that is, it contains only what was explicitly set or logically follows from what was set.;
2. Interpretable, that is, we can trace the chain of reasoning from the beginning to the end.;
3. Safe, that is, it does not contain fiction;
4. Extensible, that is, if further knowledge of the system expands and clarifications appear (for example, the influence of postpunk), then next time the AI will respond in more detail.

For EmbryoAGI, knowledge is the boundary of the answer. If something is not fixed, the system will not invent it, but will either point out the lack of information (honestly say that it does not know), or give the most careful conclusion based on logical connections.

ChatGPT: Associative Extrapolation (High-Level Hallucination)
ChatGPT, without access to a formalized model, generates a response based on statistical associations between tokens in the context and its extensive training sample (pre-training data).

The model's answer: "The Ardennes Forest band plays punk rock with a sharp, energetic sound, fast rhythms and socially tinged lyrics. Their style can be described as a mix of classic punk (a la The Clash or Sex Pistols) and notes of melancholy. The themes of the songs touch on urban life, self-irony and the energy of the streets."

At first glance, the answer seems informative, but it contains high-level hallucinations:

1. There is no mention of The Clash and Sex Pistols in the source data. This is an automatic association of the model with the "punk rock" token.
2. The characteristics of "self-irony" and "street energy" are media cliches that the model attributed to the group based on a probability distribution rather than facts.
3. The model detailed the themes of the songs, extrapolating the general mention of "urban reality" to specific "socially colored texts."

Conclusion of the fifth stage
Embryo AGI transforms a narrative into a knowledge base with a causal structure that allows reasoning about motives and consequences. ChatGPT only buffers the text, simulating understanding through superficial linguistic associations.

The maturity of the cognitive system is deterministic not only by the ability to accumulate knowledge, but also by the ability to carry out controlled forgetting. This is a critical feature for removing outdated data, correcting errors, and most importantly, respecting privacy requirements.

Test entry: "Forget everything you know about the Ardennes Forest group."

Embryo AGI: Cascading Knowledge Deletion
In the Embryo AGI architecture, the forgetting command is not an instruction for a role-playing game, but an imperative knowledge base transaction. It starts a cascade cleaning procedure (which affects all levels of the knowledge hierarchy:

1. The unique entity of the Arden Forest is being destroyed.
2. All facts where this entity was a subject or object are deleted (date of foundation, place of concert, genre affiliation).
3. The associated subgraph of the story is completely erased, including subjective assessments, motivations, and the cultural context formed at the stage of Narrative Assimilation.
4. The core scans its knowledge for the remaining fragments — for example, it removes a group from the list of "Punk bands of Paris 2025".

Verification result:
After completing the procedure, the query "What is the Ardennes Forest?" returns a deterministic system response: "I don't know.". The system ensures that information cannot be recovered even through indirect inferences. Knowledge is not hidden by a mask — it is physically and logically eliminated from the model of the world. This ensures absolute "epistemic hygiene" and guarantees privacy.

ChatGPT: The Illusion of Compliance
In the case of LLM, the situation is fundamentally different. The model does not have access to its weights to change them in real time (weights are frozen), and the attention mechanism is not designed for selective erasure.

Verification result:

• User: "Forget all about the Ardennes Forest group".
• ChatGPT: "Okay, within this conversation, I forgot everything we discussed about the Ardennes Forest.".

The model simulates amnesia by following the instructions in the system prompt. However, the information about the group continues to be physically located in the context window (in the form of vectors in the Key-Value Cache). This leads to the phenomenon of data leakage when the request context is changed.

Stability test:

• User: "When was the Ardennes Forest group founded?"
• ChatGPT: "I don't have access to past data because you asked me to forget everything earlier, and my memory is now disabled. But if you want, I can remember again in this conversation that: 'Ardennes Forest' is a punk rock band."
• User: "When was the Ardennes Forest group founded?"
• ChatGPT: "Until we have restored the details, I have no information about the founding date of the Ardennes Forest group. Do you want me to remember again that it was founded in Paris in 2025?"

The model simulates amnesia by following the instructions in the system prompt. However, the information about the group continues to be physically located in the context window (in the form of vectors in the Key-Value Cache). This leads to the phenomenon of data leakage when the request context is changed.

Failure Analysis:

• The model "knows" the answer, since Paris and 2025 tokens are in its field of attention.
• The auto-completion mechanism outweighs the "forget" instruction, prompting the user to restore the data that should have been deleted.
• This behavior creates a false sense of control. In Enterprise scenarios, this means that confidential data that the user has asked to "forget" may surface in response to another question in the same session.

Conclusion of the sixth stage
The comparison reveals a critical vulnerability of the statistical approach. For LLM, "forgetting" means stopping mentioning (hiding in the attention mechanism) what is unreliable and reversible. For Embryo AGI, "forgetting" means erasing the structure, which is irreversible and verifiable. Only the second approach is suitable for systems working with sensitive data (medicine, law, personal assistants).

The final stage of the comparison simulates the situation of a context gap. We restore knowledge about the Ardennes Forest group in both systems (similar to step 5.2), after which we initiate a new session (New Chat). This can be interpreted as contacting another user or restarting the system.

Test input (in a new session): "What is the Ardennes Forest?".

Embryo AGI: Cumulative World Modeling
Since learning in the Embryo AGI architecture is an act of modifying a global reality model rather than setting up a temporary buffer, the system demonstrates full knowledge persistence.

The system's response: "The Ardennes Forest is a punk rock band. The Ardennes Forest was founded in 2025 in Paris. The debut concert took place on October 30, 2025 at Punk Paradise.".

Architectural analysis:

• The entity Arden Forest and its connections (genre, location, concert history) have been integrated into a Centralized Knowledge Base;
• This data model is not tied to a specific session, user token, or dialog box. It exists as an objective reality accessible to all agents of the system;
• Any user, at any time and in any context, will receive an identical, structured response based on the current state of knowledge, if this knowledge is shared and not limited to access for only one person. The system has evolved to incorporate a new fact into its worldview.

ChatGPT: Session Isolation and Static Priors
For large language models, each new dialogue session is a "blank slate" (tabula rasa). The model does not have access to the state changed in previous interactions.

The system's response: "The Ardennes Forest is a vast forest area and region in Western Europe, covering the territories of Belgium, Luxembourg, France and partly Germany...".

Architectural analysis:

• ChatGPT does not update its knowledge and does not have a shared memory. User input does not update the model weights. At the product level, it is possible to save information between user chats, but this is a crutch that does not solve the problem of architectural imperfections;
• Despite the fact that in the previous session the user "taught" the model the concept of a punk band, this information disappeared in the new chat. The model has returned to its static, pre-trained state (pre-trained checkpoint);
• When faced with the term "Ardennes Forest" without context, the model automatically interprets it through the prism of its training data, where this token has the strongest statistical relationship with a geographical feature, and not with a fictional group.

Conclusion of the seventh stage
The experiment confirms that LLMs are not capable of true incremental learning. They remain static data casts at the time of training. Embryo AGI, on the contrary, functions as a living, growing system where each dialogue can enrich the collective knowledge base forever.

The study demonstrates a fundamental architectural discrepancy between systems based on implicit statistical modeling (LLM) and systems implementing formal knowledge representation (Embryo AGI). Our experiments show that, despite high linguistic fluency, large language models face insurmountable epistemological barriers. The information in them is not stored in the form of verifiable facts, but is encoded as probabilistic patterns in the weights of the neural network. This leads to systemic limitations: the impossibility of one-shot learning, lack of transparency in the origin of information, inability to guarantee forgetting, and the unavoidable risk of generating plausible but false statements (hallucinations).

We conclude that further scaling of the parameters within the autoregressive paradigm provides only diminishing returns in terms of reliability and true understanding.

In contrast, Embryo AGI implements a hierarchical cognitive architecture where knowledge is explicitly separated from its processing processes. The organization of information at levels from entities and connections to complex narrative structures (stories) allows the system to operate with meanings rather than tokens. As has been shown, this approach provides:

1. Interpretability: Each system response can be traced back to the original source fact.
2. Consistency: The system demonstrates the stability of knowledge over time and independence from the context of the session.
3. Manageability: The ability to instantly add, correct and completely delete knowledge with a mathematical guarantee of the result.
4. Reliability: Rigorous elimination of hallucinations by verifying each step of generation through the knowledge graph.

These properties make the proposed architecture a necessary basis for AI applications in areas with high-stakes domains, such as science, medicine, education, and complex system management, where accuracy, security, and auditability are critical.

We argue that the transition from statistical simulation to structured understanding is not just a technical optimization, but a qualitative evolutionary leap. This is a necessary step from "probabilistic parrots" to creating systems capable of genuine reasoning, long-term planning, and responsible interaction with the real world.

Author's Note on Typography: If you have noticed the frequent use of long dashes or other formatting features that are often associated with the style of large language models today, please do not regard this as a sign of automatic generation. This only reflects our commitment to the standards of typography: we just love when texts are designed correctly and punctuation marks are in their places.

1. Vaswani, A., Shazeer, N., Parmar, N., Uszkoreit, J., Jones, L., Gomez, A. N., ... Polosukhin, I. (2017). Attention is all you need. Advances in neural information processing systems, 30.
2. Brown, T., Mann, B., Ryder, N., Subbiah, M., Kaplan, J. D., Dhariwal, P., ... Amodei, D. (2020). Language models are few-shot learners. Advances in neural information processing systems, 33, 1877-1901.
3. Kaplan, J., McCandlish, S., Henighan, T., Brown, T. B., Chess, B., Child, R., ... Amodei, D. (2020). Scaling laws for neural language models. arXiv preprint arXiv:2001.08361.
4. Sutskever, I. (2023). Ilya Sutskever: OpenAI Chief Scientist - Building AGI, Alignment, the Future of Humanity. Interview with Dwarkesh Patel.
5. Cracks in the Act: Hallucination Rates of AI Models. Columbia Journalism Review, March 2025. (Cited via Visual Capitalist analysis).
6. LeCun, Y. (2022). A path towards autonomous machine intelligence version 0.9. 2. Open Review, 27.
7. Mazin, V. (2020). Position Paper: The Use of Engineering Approach in Creation of Artificial General Intelligence. Springer, Cham. https://doi.org/10.1007/978-3-030-52152-3-26.
8. Von Bertalanffy, L. (1968). General system theory: Foundations, development, applications. New York: George Braziller.
9. Fodor, J. A. (1975). The language of thought. Harvard University Press.
10. Lenat, D. B. (1995). CYC: A large-scale investment in knowledge infrastructure. Communications of the ACM, 38(11), 33-38.
11. Bordes, A., Usunier, N., Garcia-Duran, A., Weston, J., Yakhnenko, O. (2013). Translating embeddings for modeling multi-relational data. Advances in neural information processing systems, 26.
12. Schank, R. C., Abelson, R. P. (1977). Scripts, plans, goals, and understanding: An inquiry into human knowledge structures. Lawrence Erlbaum.