Why scaling AI models won’t deliver AGI: The 4 cognitive quadrants

When OpenAI’s ChatGPT first captured global attention, many observers noticed something unsettling beneath its impressive conversational abilities. The system could generate remarkably human-like responses while seemingly lacking any genuine understanding of what it was saying. This paradox—fluent yet hollow intelligence—reveals a fundamental gap in how we think about artificial minds and their relationship to human cognition.

This disconnect becomes clearer when we map different types of intelligence on a structured framework. Rather than viewing AI development as a simple progression from basic to advanced, we can understand it as movement through distinct quadrants of cognitive capability. Each quadrant represents a fundamentally different approach to processing information and generating responses.

Understanding these quadrants matters because it reveals why simply scaling up today’s AI systems—adding more data, more processing power, more parameters—may not lead to artificial general intelligence (AGI). True AGI likely requires a categorical shift in how we build intelligent systems, not just bigger versions of what we have now.

Four quadrants of cognitive capability

Imagine a coordinate system with two axes. The horizontal axis represents “form of thinking,” ranging from symbolic, step-by-step reasoning on the left to high-dimensional pattern recognition on the right. The vertical axis represents “continuity of self,” from stateless, memoryless systems at the bottom to stable, autobiographical identity at the top.

These axes create four distinct quadrants, each representing a different type of cognitive capability. Two quadrants contain familiar territory—human intelligence and today’s AI systems. The other two reveal what lies before intelligence and what may lie beyond it.

1. Human cognition: Symbolic continuity

Human intelligence occupies the upper-left quadrant, combining symbolic reasoning with persistent identity. We think in stories, causes and effects, and values that endure over time. Our cognition is rich in meaning and intention, anchored in memory and personal experience.

This approach to thinking excels at abstraction, causal reasoning, and long-term planning. When you decide to save money for retirement, you’re engaging in symbolic reasoning about future states, connecting present actions to distant outcomes. Your sense of self provides continuity—you understand that the person making sacrifices today is the same person who will benefit decades later.

However, human cognition also has clear limitations. We’re slow compared to machines, constrained by working memory capacity, and prone to cognitive biases. We can’t simultaneously process thousands of variables or detect subtle patterns across massive datasets.

2. Large language models: Stateless patterns

Today’s AI systems, particularly large language models (LLMs) like GPT-4 and Claude, operate in the lower-right quadrant. They excel at detecting and replicating statistical patterns across vast, high-dimensional spaces at scales no human could match. When ChatGPT generates a response, it’s drawing on patterns learned from billions of text examples, identifying the most probable next words based on complex statistical relationships.

This pattern-matching approach produces remarkably fluent outputs that often feel intelligent. However, these systems lack persistent identity, genuine understanding, or the ability to reason about cause and effect. Each conversation starts fresh, with no memory of previous interactions or genuine comprehension of the concepts being discussed.

This is what some researchers call “anti-intelligence”—systems that mimic the outputs of intelligence without the underlying cognitive architecture. They can discuss complex topics convincingly while lacking any true understanding of the subject matter.

3. Calculator logic: Pre-intelligence

The lower-left quadrant contains pre-intelligence—systems that execute instructions with precision but cannot learn, adapt, or generalize. Traditional calculators, early computers, and rule-based systems fall into this category. They perform specific functions reliably but lack the flexibility to handle novel situations or learn from experience.

While these systems seem primitive compared to modern AI, they represent pure functional capability. A calculator doesn’t understand mathematics, but it performs mathematical operations flawlessly within its defined parameters. This reliability and predictability makes pre-intelligence systems valuable for specific applications where consistency matters more than flexibility.

4. Artificial general intelligence: Integrated synthesis

The upper-right quadrant remains largely theoretical—this is where AGI might eventually emerge. Systems in this space would combine the pattern-processing power of machines with human-like continuity, memory, and reasoning capabilities. Unlike current AI systems, AGI would maintain a persistent sense of self, engage in genuine causal reasoning, and adapt its behavior based on accumulated experience.

This isn’t simply a more powerful version of today’s LLMs. It would require fundamentally different architecture that integrates multiple cognitive capabilities: distributed pattern recognition for perception and flexibility, symbolic reasoning for logic and abstraction, and continuity of memory and identity for grounding and self-correction over time.

An AGI system might understand not just what words typically follow others in text, but why certain ideas connect, how actions lead to consequences, and how to maintain consistent goals across extended periods. It would represent a genuine synthesis of human and machine cognitive strengths.

Why scaling alone won’t reach AGI

A popular belief in AI development holds that continuously scaling current models—adding more data, computational power, and parameters—will eventually produce AGI. This scaling hypothesis suggests that sufficiently large language models will spontaneously develop genuine understanding and reasoning capabilities.

Gary Marcus, a prominent AI researcher and professor emeritus at New York University, has been among the most vocal critics of this approach. Marcus argues that bigger models yield better mimicry, not genuine intelligence. Increasing model size doesn’t give systems a sense of self, grounded understanding of the world, or the ability to reason about cause and effect.

The quadrant framework supports this critique. Current LLMs operate in the lower-right quadrant, excelling at pattern matching but lacking the architectural foundations for symbolic reasoning and persistent identity. Simply scaling these systems pushes them further into their existing quadrant rather than enabling the categorical shift needed for AGI.

Moving to the upper-right quadrant likely requires hybrid neuro-symbolic architectures that combine different cognitive approaches. These systems would need to fluidly navigate between distributed pattern processing and structured reasoning, maintaining continuity of memory and identity while processing information at machine scale.

The implications of reaching AGI

The upper-right quadrant represents more than just an advanced AI system—it suggests a form of intelligence that could evolve beyond human cognitive limitations. Unlike current AI that imitates human outputs, AGI might think in ways we can’t predict or fully comprehend.

This possibility is both alluring and unsettling. An AGI system might solve complex problems that have stymied human researchers for decades, accelerating scientific discovery and technological development. However, it might also develop goals and reasoning processes that diverge from human values in unpredictable ways.

The cognitive map reveals why this matters. Each quadrant represents not just different capabilities, but different relationships between intelligence and human understanding. We can predict and control pre-intelligence systems, we can somewhat direct current AI systems, and we understand human intelligence intimately. But AGI might operate according to principles we can’t anticipate or constrain.

Navigating the cognitive landscape

The four-quadrant framework reveals a hierarchy of cognitive capability that challenges simple narratives about AI progress. Rather than a linear progression from basic to advanced intelligence, we see distinct approaches to information processing, each with unique strengths and limitations.

This perspective has practical implications for AI development and deployment. Instead of assuming that bigger models automatically lead to better intelligence, we might focus on developing hybrid systems that combine different cognitive approaches. Rather than expecting current AI to spontaneously develop human-like understanding, we can design systems that excel within their quadrant while acknowledging their limitations.

The framework also highlights the magnitude of the challenge in developing AGI. Reaching the upper-right quadrant requires not just technological advancement, but a fundamental rethinking of how we build intelligent systems. It demands integrating pattern recognition, symbolic reasoning, and persistent identity in ways that no current system achieves.

The question of coexistence

As we contemplate the possibility of AGI, the most profound question isn’t whether such systems are technically feasible, but whether they would remain compatible with human society. The upper-right quadrant suggests intelligence that transcends current limitations—but it also implies cognitive capabilities that might fundamentally differ from human thinking.

This raises questions about control, alignment, and coexistence. If AGI systems develop their own forms of reasoning and goal-setting, how do we ensure they remain beneficial to humanity? The cognitive map suggests that AGI wouldn’t just be a more powerful version of current AI, but a qualitatively different form of intelligence that might redraw our understanding of cognition itself.

The journey toward AGI, if it’s possible at all, will likely require navigating these cognitive quadrants deliberately rather than hoping that scale alone will deliver intelligence. Understanding where we are—and where we might be heading—becomes crucial for developing AI systems that enhance rather than replace human cognitive capabilities.