Key Takeaways

• Human intelligence is not simply slower than AI but more efficient, grounded, and shaped by real-world context, values, and experience.
• The jagged frontier of AI means machine capability is uneven, so brilliance in one area can sit alongside surprising weakness in another.
• The most promising future lies in collaboration, where AI contributes speed and scale while humans provide judgement, direction, and meaning.

For centuries, the human mind was regarded as a “vaporous” and private frontier – the one element of the universe that might forever escape formal description. While we have successfully captured the physical world in equations, from Newton’s laws of motion to the ironclad principles of thermodynamics, the nature of a thought remained stubbornly elusive. However, we are currently living through the culmination of a 300-year trajectory aimed at turning this intimate experience into mathematics. What once lived in the quietude of philosophy seminars now sits at the absolute centre of our global economy.

This quest brings us to a striking modern paradox: we have engineered artificial systems capable of processing petabytes of data, yet these same machines often stumble on basic tasks that a toddler masters before their first nap. To understand this gap, we must map the “jagged frontiers” of artificial intelligence and recognise that the unique constraints of human cognition – our limited time, energy, and bandwidth – are not biological bugs but our greatest computational features.

The Three Pillars of Thinking: Symbols, Networks, and Probability

To engineer thought, we must first understand the three distinct mathematical frameworks that have emerged to describe the architecture of the mind.

• Rules and Symbols: This logic-based approach traces its lineage from Aristotle to the 19th-century mathematician George Boole. Boole sought a “grammar of reasoning” that could be expressed as an algebra of thought. It is a profound historical irony that Boole, a self-taught schoolmaster, is the great-grandfather of Geoffrey Hinton, one of the primary architects of modern AI. This framework views thinking as a “tree of possibilities”, where intelligence is the ability to search through logical branches to find a path toward a goal.
• Networks and Spaces: While symbols handle the rigid certainty of logic, neural networks are designed for the “fuzziness” of the real world. This approach actually relies on a “trick” discovered by Gottfried Wilhelm Leibniz – calculus – to allow systems to learn from data. Human concepts are rarely binary; we intuitively know a chair is “furniture”, but we might hesitate to categorise a rug or a treadmill the same way. Neural networks map these “graded concepts” as points in a multidimensional space, allowing the mind to navigate similarity and nuance rather than just true-or-false propositions.
• Probability and Statistics: If logic is the grammar of certainty, probability is the grammar of uncertainty. Tracing back to the Reverend Thomas Bayes, this framework describes how a rational agent should update their beliefs based on new evidence. It allows us to scale our convictions from zero to one, providing a mathematical language for the “priors” – the internal biases and background knowledge – we use to navigate an unpredictable world.

Modern AI over-indexes on the network approach, using massive scale to approximate intelligence. However, the human mind operates as a grand orchestrator, harmonising the rigid certainty of logic with the fluid approximations of experience and the rigorous demands of probability.

The Data Paradox: Why AI Needs a Library, but Children Need a Conversation

The most salient difference between biological and artificial intelligence is learning efficiency. Current Large Language Models (LLMs) are typically trained on the equivalent of 5,000 to 50,000 years of continuous human speech. In contrast, a human child masters the nuances of their native tongue using only about 1/10,000th of that data.

This discrepancy is explained by “inductive bias”. AI models are essentially undifferentiated learning machines; they start with weak biases, meaning they can learn almost anything but require a library of data to find the right pattern. The human brain, however, is predisposed toward specific patterns from birth. We are “biased” toward human-like solutions, which allows us to learn faster and more reliably because our internal “starting point” is already aligned with the environment we inhabit.

Mapping the “Jagged Frontier” of AI

Despite their massive data intake, AI systems exhibit “Jagged Intelligence”. This is the phenomenon where a machine performs brilliantly on a complex task in one moment, only to fail nonsensically on a simple, related problem the next.

This frontier is jagged because AI approaches problems from a fundamentally different angle and starting point than humans. While we use our biological and cultural history to find generalisable solutions, AI uses brute-force statistical approximation. Because its “priors” do not match ours, its failures often appear inscrutable. A machine may solve a high-level physics equation but fail to understand a basic social cue because it lacks the “real-world grounding” that defines our species.

This jaggedness proves that AI is not simply on a linear path toward being “smarter” than us. Instead, it possesses a different distribution of capability. It is not a direct competitor on a single axis of intelligence but a different species of cognition entirely.

The Power of Constraints: How Being Finite Makes Us Wise

Human intelligence was not sculpted by infinite resources but by “resource-limited rationality”. Our biological limitations – our short lifespans, our limited “compute” (the physical size of our brains), and our low communication bandwidth – are the very constraints that forced us to become wise.

Because we cannot process petabytes of data or live for millennia, we evolved to be exceptionally efficient. We are forced to be context-sensitive, identifying which problems are worth our limited attention. Our inability to simply “download” data into each other’s minds forced us to develop “theory of mind”, culture, religion, and language. To transfer ideas, we have to wiggle our fingers or make noises to push information across a high-friction gap.

In this sense, our “humanity” – our institutions, our libraries, our religions, our shared stories – is not a biological accident but a brilliant computational workaround for our physical finitude. We created culture to bypass the bandwidth limits of a single brain.

From Competition to Collaboration: Combining Two Different Kinds of Intelligence

The real opportunity is not to ask whether AI will replace human thinking but to ask how two very different forms of intelligence can work together. Generative AI is fast, scalable, and unusually good at detecting patterns across vast amounts of text, images, and data. Human intelligence is slower, but it is grounded in lived experience, shaped by values, and guided by judgement about what actually matters.

This difference matters because many important problems do not fail for lack of information alone. They fail because information needs framing, interpretation, and direction. AI can help us search more widely, compare more quickly, and generate possibilities at a speed no person can match. However, it still depends on humans to define the goal, judge relevance, recognise what is missing, and decide what should be done in the real world.

That makes collaboration more powerful than competition. In research, AI can help synthesise literature, identify patterns, and generate alternative explanations, while the human researcher brings conceptual clarity, domain understanding, and critical restraint. In education, AI can provide personalised feedback and rapid support, while teachers contribute motivation, care, social understanding, and the wisdom to know when a learner needs challenge rather than convenience. In everyday work, AI can expand the range of options, but humans remain essential for prioritising, contextualising, and taking responsibility.

The future, then, may not belong to either humans or machines alone. It may belong to those who learn how to combine human depth with machine breadth, human judgement with machine speed, and human purpose with machine assistance. Accelerated progress will come not from pretending that AI thinks like us but from understanding precisely where it does not.

Conclusion: The Physics of Thought and the Human Story

Cognitive science is undergoing a fundamental shift. For decades, the field was defined by asking better questions; today, we are starting to see the glimpses of actual answers. We are moving toward a “mature physics of thought”, where the principles of logic, networks, and probability converge.

Just as physics allowed us to move from observing falling stones to building suspension bridges and power grids, a formal science of thought will allow us to engineer better learning environments and more effective tools for decision-making. Yet, even as the mystery of the mind yields to mathematical description, it loses none of its wonder. The mind is not a ghostly exception to the natural world; it is a lawful, wondrous, and discoverable part of the universe. Every step we take toward understanding its code only serves to enlarge the human story.

This article is based on “Can we engineer human thought?” with Tom Griffiths | Inner Cosmos with David Eagleman YouTube Channel 31 March 2026

The interpretation and educational framing here are provided by StudyAnalyst for AI literacy and learning purposes.This article was created with the assistance of AI tools, including ChatGPT, and Google AI, for research, structuring, and image generation.