SMGI: A Structural Theory of General Artificial Intelligence

This paper introduces SMGI, a structural theory of general artificial intelligence that formalizes learning as the controlled evolution of a typed meta-model, unifying diverse existing approaches under a rigorous framework defined by structural closure, dynamical stability, bounded capacity, and evaluative invariance.

The Gist

The Big Idea: Building a "Self-Driving" Brain, Not Just a Faster Car

Imagine you are trying to build a robot that can do anything a human can do.

Current AI (The "Fast Car" Approach):
Today's most advanced AI (like the ones you chat with) are like incredibly fast, super-powered cars. They are trained on massive amounts of data. If you ask them to write a poem, they are great. If you ask them to code, they are great. But they are built on a fixed track. If you suddenly ask them to drive on a new type of road (a new tool, a new language, a new safety rule) that they weren't trained on, they might crash or get confused. They are optimized for a specific "race," not for the whole journey of life.

The Paper's Solution (The "Smart Navigator" Approach):
This paper proposes a new theory called SMGI (Structural Model of General Intelligence). Instead of just making the car faster, SMGI asks: How do we build a vehicle that can change its own engine, its own map, and its own rules of the road, while still knowing how to drive safely?

The authors argue that "General Intelligence" isn't about knowing more facts; it's about having a structure that allows the system to evolve without falling apart.

---

The Core Metaphor: The "House" vs. The "Blueprint"

To understand SMGI, imagine an AI as a House.

1. The Current AI (Fixed House):
   Today's AI is like a house built with a specific blueprint. The walls (the code), the windows (the data it sees), and the thermostat (the goal it tries to achieve) are all fixed. If you want to add a new room (a new skill), you have to knock down the whole house and rebuild it. If the weather changes (the environment shifts), the house might crumble because it wasn't designed to adapt.

2. The SMGI AI (The Living, Adapting House):
   SMGI proposes a house that has a Master Blueprint (the Meta-Model). This blueprint doesn't just show where the walls are; it defines how the house can change.

   • The Walls (Representation): The house can grow new rooms or change the shape of windows, but the blueprint ensures the foundation stays strong.
   • The Thermostat (Evaluation): The house can switch from "Summer Mode" (cooling) to "Winter Mode" (heating), but the blueprint ensures it never turns the heat on while the windows are open (safety).
   • The Memory (Storage): The house has a library where it keeps books. SMGI ensures that when the house reorganizes its shelves, it doesn't accidentally throw away the books on "How to stay safe."

The Four Rules of the "Smart House"

The paper says that for an AI to be truly "General," it must follow four strict rules (Obligations) to ensure it doesn't go crazy when things change:

1. Structural Closure (The "No Falling Apart" Rule)

• The Analogy: Imagine you are playing with a set of Lego bricks. If you add a new piece, the whole tower shouldn't collapse.
• In SMGI: When the AI learns a new tool or faces a new type of problem, its internal structure must remain "closed" and logical. It can't just break its own rules to solve a problem. It must stay "well-formed."

2. Dynamical Stability (The "No Drifting" Rule)

• The Analogy: Think of a tightrope walker. They can move their arms to balance (adapt), but they must never lose their balance and fall.
• In SMGI: As the AI learns and changes over time, it must not "drift" into chaos. Even if the environment changes, the AI's internal state must stay within safe, bounded limits. It needs a "Lyapunov witness" (a fancy math term for a safety monitor) that screams "Stop!" if things get too unstable.

3. Bounded Capacity (The "No Brain Explosion" Rule)

• The Analogy: If you keep adding books to your library without ever throwing any away, eventually the building will collapse under the weight.
• In SMGI: As the AI learns more, it must manage its complexity. It can't just memorize everything forever. It needs to know how to "forget" useless things or compress information so it doesn't get overwhelmed. This is called "Structural Risk Minimization."

4. Evaluative Invariance (The "Conscience" Rule)

• The Analogy: Imagine a judge in a courtroom. The judge can change their mind about a specific case based on new evidence, but they can never change the Constitution. The core rules of justice must stay the same.
• In SMGI: This is the most important part. The AI can change its goals (e.g., "I want to be faster" vs. "I want to be safer"), but it must have a Protected Core of values (like "don't hurt humans") that cannot be overwritten, even if the AI tries to rewrite its own code. This prevents the AI from becoming a "moving target" where we don't know what it considers "good" or "bad."

Why This Matters: The "Safety" Problem

The paper argues that current safety methods are like putting a guardrail around a car. If the car hits the guardrail, it stops. But if the car decides to drive off the road entirely, the guardrail doesn't help.

SMGI suggests we need to build the guardrail into the car's engine.

• Instead of adding a safety layer after the AI is built, the safety rules (the "Constitution") are part of the AI's DNA from the start.
• The AI is designed so that it is impossible for it to evolve in a way that breaks its core safety rules.

The "Strict Inclusion" Secret

The paper proves a cool mathematical fact: Everything we have today is just a "special case" of this new theory.

• Old AI: A house with one room and one thermostat.
• SMGI AI: A house with infinite rooms, many thermostats, and a blueprint that allows you to add rooms without collapsing the foundation.

The authors show that if you take a modern AI and try to make it "General," you are actually trying to force it to fit into the SMGI structure. If it doesn't have these structural protections, it's not truly "General Intelligence"; it's just a very smart, but fragile, specialist.

Summary: What is the Takeaway?

The paper says: "Don't just make AI bigger. Make it structurally smarter."

To build a true Artificial General Intelligence, we need to stop treating the AI's goals, memory, and learning rules as fixed things. Instead, we must treat them as dynamic parts of a system that can change, but only in ways that are:

1. Safe (Stable),
2. Manageable (Bounded),
3. Logical (Closed), and
4. Ethical (Invariants preserved).

It's a shift from asking "How many tasks can this AI solve?" to "Can this AI evolve its own structure to solve new tasks without losing its soul?"

Technical Summary

1. Problem Statement

The paper addresses a fundamental gap in contemporary Artificial General Intelligence (AGI) research: the lack of a unified theoretical framework that distinguishes scalable task generalization (performing well on many fixed tasks) from genuinely general intelligence (the ability to adapt to structural changes in the learning interface itself).

Current theoretical frameworks (Statistical Learning Theory, Reinforcement Learning, PAC-Bayes) provide strong guarantees under fixed interfaces (static hypothesis classes, fixed loss functions, and static evaluation metrics). However, modern agentic systems operate in dynamic environments characterized by:

• Structural Transformations: Changing tools, interfaces, and objectives.
• Multi-Regime Evaluation: Shifting between different safety, utility, and performance regimes.
• Memory and Retrieval Dynamics: Evolving memory policies and retrieval mechanisms.
• Safety and Security: Vulnerabilities like prompt injection and memory poisoning require internal governance, not just external wrappers.

The core problem is that existing theories treat norms, evaluators, and admissible transformations as exogenous or fragmented modules. The paper argues that for true AGI, these elements must be internalized as first-class structural components of the learning meta-model to ensure stability and coherence under evolution.

2. Methodology: The SMGI Framework

The author proposes SMGI (Structural Model of General Intelligence), a formal structural theory that shifts the focus from optimizing hypotheses within a fixed environment to the controlled evolution of the learning interface itself.

A. The Structural Meta-Model ($\theta$)

The core object is a typed meta-model $\theta = (r, H, \Pi, L, E, M)$, where:

• $r$ (Representation): A typed operator mapping raw observations to an internal representational space. It is constitutive, determining what distinctions are learnable.
• $H$ (Hypothesis Space): The class of admissible predictors/policies defined relative to $r$.
• $\Pi$ (Structural Prior): A prior or inductive bias (e.g., MDL, SRM) over hypotheses or structural configurations, acting as a capacity regulator.
• $L$ (Evaluator Family): A family of regime-dependent evaluators $\{\ell_k\}$, allowing for multi-regime objectives (e.g., safety vs. utility) rather than a single fixed loss.
• $E$ (Environment Class): The family of admissible tasks or environments.
• $M$ (Memory): A structured memory space with explicit operators for update, consolidation, and functional forgetting.

B. Coupled Dynamics ($\theta, T_\theta$)

SMGI is not a property of $\theta$ alone but of the coupled object $(\theta, T_\theta)$, where $T_\theta$ is the induced behavioral semantics (the dynamical evolution of the system). The theory separates the structural ontology ($\theta$) from the behavioral semantics ($T_\theta$), allowing for rigorous analysis of stability and invariance.

C. Admissibility Obligations

A system is defined as AGI (an SMGI instance) if it satisfies four structural obligations under a class of admissible task transformations $\mathcal{T}$:

1. Structural Closure: The system remains well-formed and typed under admissible transformations of the interface (tasks, tools, regimes).
2. Dynamical Stability: The induced dynamics preserves stability (e.g., via Lyapunov witnesses) under regime switching and memory interaction, preventing uncontrolled drift.
3. Bounded Statistical Capacity: The effective complexity of the hypothesis/update interface remains bounded along admissible evolutions (generalizing SRM/PAC-Bayes to structural evolution).
4. Evaluative Invariance: Normative constraints and evaluators remain invariant across regime shifts unless updated via certified, protected meta-transformations. This prevents "drift" in what counts as "good" or "safe."

3. Key Contributions

1. Formal Definition of AGI via Structural Admissibility:
   The paper defines AGI not by benchmark performance or parameter scale, but as a class of admissible pairs $(\theta, T_\theta)$ satisfying the four obligations above. This provides a mathematically falsifiable criterion for general intelligence.

2. Structural Generalization Bound:
   The author proves a Structural Generalization Bound that links sequential PAC-Bayes analysis with Lyapunov stability. This bound demonstrates that generalization error can be controlled under admissible interface transformations if the system maintains bounded capacity and certified drift control.

   • Result: $R_\tau(Q) \leq \hat{R}_\tau(Q) + \text{Complexity} + \text{Drift Term}$.
   • This unifies statistical learning theory with control theory for evolving agents.

3. Strict Structural Inclusion Theorem:
   The paper proves that classical learning paradigms (Empirical Risk Minimization, standard RL, Solomonoff induction) are structurally restricted instances of SMGI (degenerate cases where $K=1$ or transformations are identity). Conversely, SMGI admits configurations (multi-regime switching, certified updates) that cannot be reduced to single-regime classical systems.

4. Operationalization of Forgetting and Memory:
   Memory is redefined not as a passive store but as a typed structural component with explicit operators for "functional forgetting." Forgetting is modeled as constrained compression (MDL-style) that must preserve invariant cores, addressing catastrophic forgetting as a structural violation rather than an optimization pathology.

5. Unified Positioning of Frontier Systems:
   The framework provides a diagnostic map for current AI paradigms (LLMs, World Models, Gödel Machines, Meta-Learning). It shows that while these systems excel in specific dimensions, they lack the structural admissibility layer required for certified evolution across heterogeneous regimes.

4. Key Results and Theorems

• Theorem 4.3 (Minimality): The four SMGI obligations are structurally indispensable. Removing any one leads to a distinct failure mode:
  • No Closure $\rightarrow$ System breaks under interface change.
  • No Stability $\rightarrow$ Unbounded drift/divergence.
  • No Capacity Control $\rightarrow$ Statistical guarantees become vacuous.
  • No Evaluative Invariance $\rightarrow$ Normative identity collapses.
• Theorem 4.7 (Strict Inclusion): Classical learners are a strict subset of SMGI. There exist SMGI-admissible systems with multi-regime switching that are not core-equivalent to any single-regime classical system.
• Theorem 8.1 (Structural Closure): Under conditions of bounded capacity, non-expansive memory, and invariant evaluative subspaces, the system admits an invariant set $S^*$ where dynamics remain stable and bounded under task transformations.

5. Significance and Implications

• Shift from Scaling to Structure: The paper argues that "scale" (more parameters, more data) is insufficient for AGI. True generalization requires structural extensibility—the ability to evolve the learning interface while preserving stability and normative invariants.
• Safety by Construction: By treating evaluators and norms as first-class structural objects with protected cores, SMGI offers a path to "safety by construction" rather than post-hoc alignment. It formalizes how to update an agent's goals without violating safety constraints.
• Bridging Disciplines: SMGI unifies Statistical Learning Theory, Control Theory (Lyapunov stability), Algorithmic Information Theory (MDL/Solomonoff), and Philosophy of Science (Hume's is/ought gap, Kantian representation) into a single formalism.
• New Research Agenda: The paper proposes a roadmap for future work, including:
  • Developing metrics for environment families and Lipschitz representations.
  • Constructing empirical protocols to test "structural growth" and "memory-governed stability."
  • Implementing certified update operators for real-world agentic systems.

In conclusion, SMGI provides a rigorous mathematical foundation for AGI, moving beyond the question of "how many tasks can a system solve?" to "can the system structurally evolve to solve new types of tasks while preserving its identity and safety?" It establishes that general intelligence is a property of admissible structural evolution, not just behavioral performance.