An exploration of whether intelligence can be reduced to computation, whether thinking can be represented mathematically, and what it means to build an intelligent machine.
Introduction: The Question Behind Artificial Intelligence
We have built machines that can calculate faster than any human, remember more information than any individual could possibly retain, recognise patterns across enormous datasets, translate languages, generate paintings, compose music, and engage in conversations that sometimes feel remarkably human.
Yet, a fundamental question remains:
Can intelligence itself be computed?
It is an interesting question because it forces us to look beyond artificial intelligence and examine intelligence in its original form: biological intelligence.
When a human solves a mathematical problem, recognises a familiar face, understands a metaphor, learns a new language, or makes a decision under uncertainty, what exactly is happening inside the mind? Are these activities fundamentally computational processes? If so, can we describe them mathematically and reproduce them using machines? Or is there something about intelligence that cannot be captured by computation alone?
To explore this, we must first understand what we mean by computation.
What Does It Mean to Compute Something?
At its simplest, computation is the process of transforming information according to a set of rules.
Consider a simple mathematical operation:
We provide an input, apply an operation, and receive an output. The same principle applies to a computer program. A computer receives information, follows instructions, performs operations, and produces a result.
But computation is not limited to arithmetic. Computers can process text, images, sound, video, and complex mathematical structures. They can identify relationships between data, execute logical operations, simulate physical systems, and learn statistical patterns.
In computer science, computation can be understood as a transformation from one state of information into another:
Here, represents the current state, represents new input, represents the operation or transformation, and represents the resulting state.
This simple abstraction becomes surprisingly powerful when applied to intelligent behaviour.
Imagine a system receiving visual information, comparing it with previously acquired information, identifying an object, and deciding what action to take. The system continuously transforms information from one state into another.
Now consider a human being doing the same thing.
You see a glass on a table. You recognise it as a glass, understand that it can contain water, estimate its position, reach towards it, and pick it up. Could this entire process be described as computation?
Perhaps. But the complexity lies in understanding what happens between receiving information and producing an intelligent response.
The Human Brain: A Biological Information-Processing System
The human brain contains approximately 86 billion neurons, connected through an extraordinarily complex network of synapses.
Neurons communicate through electrical and chemical signals. They receive information, integrate signals, and influence the activity of other neurons. These interactions form the biological foundation of perception, memory, learning, reasoning, and other cognitive processes.
At a very simplified level, a neuron can be represented mathematically:
Here, represents incoming signals, represents the strength of each connection, represents a bias, represents an activation function, and represents the resulting output.
Does this look familiar? It should.
This is also a fundamental mathematical representation used in artificial neural networks.
The resemblance is not accidental. Artificial neural networks were inspired, in part, by biological nervous systems. However, an artificial neuron is an extremely simplified mathematical abstraction of a biological neuron. Real neurons involve complex biochemical processes, different electrical behaviours, temporal dynamics, and interactions with their surrounding biological environment.
Nevertheless, the mathematical abstraction provides an important starting point.
The brain receives sensory information, processes it, stores and modifies information, learns from experience, and generates responses. These activities suggest that many aspects of human cognition can be investigated as information-processing mechanisms.
But there is an important distinction: showing that the brain processes information does not automatically prove that every aspect of intelligence is computational.
That is where the philosophical and scientific questions begin.
The Computational Theory of Mind
One influential idea in cognitive science and philosophy is the Computational Theory of Mind. Its central proposition is that the mind can be understood, at least in significant part, as a computational system.
According to this perspective, mental processes involve the manipulation of representations according to certain rules.
Consider the sentence: If it rains, the ground becomes wet.
Now suppose we observe that it is raining. We can logically infer that the ground may become wet. The mind appears to manipulate representations and establish relationships between them.
This resembles a computational process. We have an existing piece of knowledge, receive new information, establish a relationship between them, and derive a conclusion.
Traditional artificial intelligence explored this idea through symbolic reasoning. Knowledge could be represented using symbols, rules, logical relationships, and structured representations. A system could then manipulate those representations to solve problems.
Chess engines can evaluate possible moves. Expert systems can apply predefined rules to reach conclusions. Logic-based systems can derive new information from existing facts. Planning systems can search for sequences of actions that achieve a goal.
These are examples of computational systems performing tasks associated with intelligent behaviour.
However, human intelligence does not always operate through explicit logical rules.
We recognise faces without consciously calculating facial measurements. We understand familiar languages without explicitly applying grammatical rules to every sentence. We often make decisions based on intuition before we can explain our reasoning.
This raises another question: if intelligence is computation, must all computation involve explicit symbols and logical rules?
Not necessarily.
From Symbolic Reasoning to Neural Computation
There are different ways to approach computation, and these approaches have shaped two important traditions in artificial intelligence: symbolic AI and connectionist AI.
Symbolic AI attempts to represent knowledge explicitly. Imagine teaching a machine to identify a bird. We might define rules stating that an object with wings, a beak, and feathers may be a bird. The machine applies these rules to the available information.
This approach is useful when knowledge, rules, and reasoning need to be explicitly represented. However, the real world is rarely as neatly structured as our rules. What happens when the object is partially hidden? What if it is a toy bird? What if the machine encounters a species it has never seen before?
Writing rules for every possible situation quickly becomes impractical.
Connectionist approaches take a different direction. Instead of explicitly programming every characteristic of a bird, we can provide a neural network with thousands of labelled images. During training, the network adjusts its internal parameters to improve its predictions. Eventually, it may learn to recognise birds in images it has never encountered before.
This is the foundation of modern deep learning.
The distinction is important. Symbolic systems primarily manipulate explicitly represented knowledge, whereas neural networks learn statistical representations from data.
Modern AI increasingly combines these approaches. A system might use a neural network to recognise objects and a symbolic reasoning mechanism to plan actions based on those objects.
Intelligence may therefore involve several complementary computational mechanisms rather than one universal algorithm.
Can Learning Be Computed?
Consider how a child learns to recognise a dog.
Initially, the child may not understand what distinguishes a dog from other animals. But after seeing different dogs, hearing people use the word, observing their behaviour, and interacting with them, the child gradually develops an understanding.
Learning changes the internal state of the learner.
In machine learning, we can describe a similar process. A model receives training data, generates predictions, measures the difference between its predictions and the expected results, and adjusts its parameters.
A simplified learning process can be represented as:
Training Data → Model Prediction → Calculate Error → Adjust Parameters → Improved Model
Mathematically, suppose a model has parameters represented by . We want to minimise a loss function:
Here, represents the model parameters, represents the error or loss, and represents the parameters that minimise that loss.
This is a fundamental idea behind many machine learning algorithms. The model learns by adjusting its internal parameters based on experience.
But human learning is much more than minimising a mathematical loss function.
Humans learn through observation, imitation, experimentation, language, emotions, social interaction, curiosity, and physical experience. We can learn from very few examples, transfer knowledge from one domain to another, and recognise that an entirely new problem resembles something we encountered years ago.
Modern AI systems are beginning to demonstrate some of these capabilities, but they still differ significantly from human learning in important ways.
The question is not simply whether learning can be computed. It is whether the computational mechanisms we currently understand are sufficient to explain the remarkable flexibility of biological learning.
Is Intelligence Just Pattern Recognition?
A considerable part of modern AI is built around identifying patterns.
A language model identifies statistical relationships between tokens. An image recognition model identifies patterns in pixels. A recommendation system identifies patterns in user behaviour.
Pattern recognition is incredibly powerful. But is intelligence merely pattern recognition?
Consider an example. You see dark clouds gathering in the sky. You remember that similar clouds have previously been followed by rain. You conclude that it might rain soon.
This involves pattern recognition.
But now imagine that you have never seen rain. Someone explains the concept of rain to you. You understand that water droplets can fall from clouds, make surfaces wet, and affect outdoor activities. You can reason about situations you have never personally experienced.
This introduces several additional capabilities.
Generalisation allows us to apply learned knowledge to unfamiliar situations. Abstraction allows us to identify common principles across different experiences. Causal reasoning helps us understand why something happens. Analogy allows us to transfer knowledge between seemingly unrelated situations. Counterfactual reasoning enables us to imagine what might happen under different circumstances.
These capabilities are essential to many forms of intelligent behaviour.
Pattern recognition helps us identify what resembles something we already know. Reasoning helps us explore what might be true beyond what we have directly observed.
Modern machine learning systems can demonstrate aspects of all these capabilities, although their reliability and depth vary considerably.
Understanding the relationship between pattern recognition, reasoning, and intelligence remains one of the central challenges in AI research.
The Limits of Computation: The Question of Consciousness
So far, we have explored intelligence through information processing, learning, reasoning, and problem-solving.
But there is another dimension to the discussion: consciousness.
Imagine two systems.
The first system receives information, recognises an object, classifies it, and produces an appropriate response.
The second system does exactly the same thing, but also has a subjective experience of seeing that object.
Is there a difference?
From the outside, perhaps we cannot determine it.
The first system might behave intelligently without any subjective experience. The second might possess an experience of seeing, understanding, or feeling.
This introduces what philosopher David Chalmers famously called the hard problem of consciousness.
The hard problem asks why and how physical processes give rise to subjective experience.
Why does processing information feel like something? Why is there an experience of seeing red rather than merely processing information about a particular wavelength of light? Why does pain hurt? Why does music evoke an emotional experience?
These questions are different from asking how the brain processes sensory information.
We can investigate the neural mechanisms involved in vision, attention, memory, and emotion. But explaining those mechanisms does not necessarily resolve why subjective experience exists.
There are several philosophical positions on this subject.
Physicalism holds, broadly, that consciousness is grounded in the physical world. Functionalism focuses on the functional organisation of mental states rather than the particular material from which a system is constructed. Computational theories of mind explore whether mental processes can be understood through computation.
Other perspectives question whether computation or physical descriptions alone can adequately explain subjective experience.
There is no universally accepted solution.
Importantly, demonstrating intelligent behaviour in an AI system does not establish that it is conscious. Likewise, the question of whether a sufficiently advanced computational system could possess consciousness remains open.
We should be careful not to confuse intelligence, the ability to behave intelligently, and subjective experience. They are related questions, but they are not interchangeable.
Is the Universe Itself Computational?
There is an even more interesting direction we can take this discussion.
What if computation is not merely something humans invented to build computers? What if computation is connected to the fundamental structure of the physical universe?
Physics describes the universe through mathematical relationships. Particles interact. Physical systems evolve. Information can be stored, transformed, and transmitted.
Some theories in physics and computer science explore whether physical processes can be understood as forms of computation.
The ideas of digital physics and the computational universe suggest that the physical world might, at a fundamental level, have a computational description. Researchers have explored connections between computation, information theory, quantum mechanics, and the laws of physics.
However, this is not an established conclusion that the universe literally operates like a conventional computer. It is an ongoing area of theoretical exploration.
Why does this matter to artificial intelligence?
Because if intelligence emerges from physical processes, and physical processes can be modelled computationally, perhaps intelligence can also be reproduced through computation.
But there is a distinction between modelling a phenomenon and reproducing everything about that phenomenon.
A weather simulation can predict atmospheric behaviour without creating an actual atmosphere. A computational model of a brain might reproduce certain cognitive functions without necessarily reproducing biological life or subjective experience.
The relationship between simulation, implementation, and reality deserves careful consideration.
What Does This Mean for Artificial General Intelligence?
Artificial General Intelligence, or AGI, is often discussed as the possibility of building machines capable of performing a broad range of intellectual tasks rather than being restricted to a narrow set of specialised functions.
But what would it actually take to achieve it?
Would a machine need to learn continuously from its environment, develop and apply abstract concepts, reason about unfamiliar problems, understand cause-and-effect relationships, plan and adapt its actions, transfer knowledge across different domains, understand its own limitations, and operate effectively in the physical world?
These are not necessarily independent abilities. They interact with one another to produce flexible intelligence.
A machine might be excellent at language generation but struggle with physical reasoning. Another might perform mathematical calculations exceptionally well but lack the ability to adapt to unfamiliar environments.
Human intelligence itself is not a single isolated ability. It involves an interconnected collection of cognitive capabilities.
This suggests that building general intelligence may require more than scaling a single computational architecture. It may require combining learning, memory, reasoning, perception, planning, action, and adaptation into a coherent system.
Whether such a system would also possess consciousness is a separate question.
We do not yet have a complete scientific understanding of human intelligence, let alone a definitive computational theory that explains every aspect of it.
Nevertheless, the progress of AI demonstrates that increasingly sophisticated intelligent behaviour can emerge from computational systems.
That alone makes the question worth pursuing.
So, Can Intelligence Be Computed?
Perhaps the most honest answer is that we do not know the complete answer yet.
We have strong evidence that many components of intelligence can be computationally modelled. We can build systems that learn, recognise patterns, solve problems, generate language, make predictions, and perform increasingly complex tasks.
We also know that biological intelligence emerges from physical processes involving extraordinarily complex networks of neurons and other biological mechanisms.
These observations provide a strong foundation for investigating intelligence computationally.
But several questions remain unresolved.
Can every aspect of human intelligence be represented computationally? Can a machine develop the same flexibility, adaptability, and understanding that humans demonstrate? Can subjective experience emerge from computation? Are there fundamental limits to what computation can achieve?
These are scientific and philosophical questions that require different kinds of investigation.
Perhaps the mistake would be to assume that intelligence must either be entirely computational or entirely beyond computation.
It is possible that computation explains many of the mechanisms underlying intelligence while other questions require deeper theories of biology, physics, and consciousness.
And perhaps our current understanding of computation itself is still incomplete.
A Thought to End With
When we look at a human being, we see more than a biological machine.
We see memories, emotions, imagination, curiosity, creativity, reasoning, and the ability to question our own existence.
Yet all of these experiences are somehow connected to a physical brain.
We are trying to understand how matter becomes capable of thinking about matter itself.
And now, through artificial intelligence, we are attempting to recreate some of those capabilities using a different physical medium.
There is something remarkable about this endeavour.
For the first time in history, intelligence is not only trying to understand the world. It is also trying to understand itself by building another form of intelligence.
Perhaps the journey towards artificial general intelligence will teach us as much about ourselves as it teaches us about machines.
And perhaps the most important question is not whether we can build something that thinks.
It is whether, in trying to build it, we will finally understand what thinking really means.
This is the sixth exploration in my Artificial Intelligence series. The objective is not simply to learn how AI works, but to understand the deeper ideas behind intelligence, computation, learning, and the human mind.