Can Computer Scientists Build a Brain? A Deep Analysis of Two Technical Approaches

An Ancient Question Spanning Multiple Disciplines

"Can computer scientists build a brain?" — this seemingly simple question actually touches upon one of the most profound debates at the intersection of artificial intelligence, neuroscience, and philosophy. When we talk about "building a brain," what exactly do we mean? Are we talking about replicating the physical structure of the human brain's approximately 86 billion neurons, or merely reproducing its functions and intelligence? The difference between these two paths determines the entire direction of the discussion.

With the breakthroughs of large language models (LLMs) in reasoning, generation, and multimodal understanding, more and more people are asking: Have these systems already "approached" the brain in some sense? Or does an insurmountable chasm still separate them from true biological intelligence?

Two Technical Approaches to "Building a Brain"

Bottom-Up: Simulating Neurons to Construct an Artificial Brain

The first approach starts from biology, simulating the brain's physical architecture layer by layer. The EU invested over one billion euros in the Human Brain Project (2013-2023), attempting to simulate the brain's neural circuits using supercomputers. This ambitious project was initially led by neuroscientist Henry Markram, with the goal of completely simulating the human brain on supercomputers. However, the project underwent significant directional adjustments during its execution: shifting from full-brain simulation to building digital infrastructure platforms for neuroscience research (such as EBRAINS). Its earlier phase, the Blue Brain Project, successfully simulated a cortical column of the rat brain (approximately 31,000 neurons), but scaling this to the complete human brain requires computational resources and data precision far beyond current technological capabilities. Supporters of this approach believe that only by truly replicating neurons, synapses, and their dynamic connections can we capture the essence of intelligence.

However, the real-world challenges are extremely daunting. The complexity of the human brain far exceeds any existing computational system:

• Each neuron may connect to thousands of other neurons, with the total number of synaptic connections in the brain estimated at approximately 100 trillion
• Synaptic plasticity and the chemical dynamics of neurotransmitters are difficult to model precisely
• The brain's energy efficiency is only about 20 watts, far lower than any computer of equivalent computational power

Even if we could precisely scan a brain's connectome, we might not understand how it works — just as having a complete city map doesn't mean understanding how the city functions. Progress in connectomics provides an intuitive reference for this difficulty: the adult fruit fly whole-brain connectome published in 2024 is one of the field's milestones, but even for this brain with only about 140,000 neurons, complete mapping took years of time and enormous resources. The leap from fruit fly to human brain spans several orders of magnitude.

Top-Down: The Triumph of Functionalist AI

The second approach is the path currently taken by mainstream AI: not caring about how the brain implements things, but only whether its functions can be reproduced. The "neural networks" in deep learning borrow vocabulary from neuroscience, but their mathematical essence is far removed from real neurons. The Backpropagation algorithm — the core mechanism of deep learning training, formally popularized by Rumelhart, Hinton, and Williams in 1986 — is almost certainly impossible in biological brains. This is because backpropagation requires perfectly symmetric weights for forward and backward paths (the "weight transport problem"), and no known mechanism in the real brain supports this; furthermore, biological neurons transmit signals through discrete action potentials (spikes) rather than the continuous gradient values that backpropagation relies on. In recent years, researchers have proposed various biologically plausible alternatives, including Hinton's Forward-Forward Algorithm proposed in 2022, Predictive Coding theory, and local learning methods based on Hebbian learning rules, attempting to bridge the learning mechanism gap between artificial and biological neural networks.

The Transformer architecture is an even purer engineering product. It was proposed by a Google team in their 2017 paper Attention Is All You Need, originally designed to solve sequence modeling problems in machine translation. Its core innovation — the Self-Attention mechanism — allows the model to attend to information from all other elements in a sequence when processing each element, thereby capturing long-range dependencies. From a computational perspective, self-attention is essentially a weighted sum of the input sequence, with weights dynamically generated through Query-Key-Value matrix operations. Although some researchers have tried to find neural circuits analogous to attention mechanisms in the biological brain (such as thalamic gating of cortical signals), the Transformer's global attention computation is fundamentally different from the brain's local connectivity patterns — it is entirely the result of engineering optimization, not an imitation of biology.

The philosophical foundation of this approach is "Functionalism," proposed by Hilary Putnam in the 1960s and later developed by Jerry Fodor and others. The core claim of functionalism is that mental states are defined by their functional roles — that is, their causal relationships with inputs, outputs, and other mental states — rather than by their physical implementation. This means that the same mental state can in principle be realized by carbon-based neurons, silicon-based chips, or any other physical medium, as long as the functional relationships remain consistent (i.e., "multiple realizability"). If a system can exhibit intelligent behavior, then it is functionally equivalent to intelligence, regardless of its physical implementation. The success of the GPT series of models has, to some extent, validated the viability of this path — we don't need to understand the brain's biological mechanisms to build machines that can "think."

However, functionalism also faces serious philosophical challenges: John Searle's Chinese Room argument questions whether pure functional simulation can produce genuine understanding; Ned Block's "China Brain" thought experiment questions the reasonableness of infinitely extending functional equivalence; and Thomas Nagel's "What Is It Like to Be a Bat" points out that functionalism may be unable to explain the nature of subjective experience. These challenges have not been decisively answered, and they hint at the deep limitations of the functionalist approach.

Intelligence and Consciousness: The Core Divide of the Artificial Brain

Even if computer scientists successfully build a system functionally equivalent to the brain, a thornier question remains unresolved: Would it possess consciousness?

Intelligence and consciousness are two fundamentally different concepts:

• Intelligence concerns the ability to solve problems
• Consciousness concerns subjective experience — what philosophers call "qualia"

Qualia refer to the subjective qualitative character of conscious experiences — such as the "redness" when seeing red, the taste experience when eating chocolate, or the auditory quality when hearing a chord. Philosopher David Chalmers drew an influential distinction in 1995: the "easy problems" of consciousness involve explaining the neural mechanisms of cognitive functions (such as attention, memory, behavioral control), which can in principle be solved through computational and neuroscientific methods; while the "Hard Problem of Consciousness" asks why these physical processes are accompanied by subjective experience — why isn't information processing done "in the dark," but instead carries a quality of "what it feels like"?

A system can surpass humans at Go, programming, and writing, yet may not "feel" anything at all. This is precisely the most easily confused point in current AI discussions.

Current major theories of consciousness give different predictions about whether AI can possess consciousness. Giulio Tononi's Integrated Information Theory (IIT) holds that consciousness correlates with the degree to which a system integrates information (quantified by the Φ value), suggesting that current AI systems based on feedforward architectures may have extremely low Φ values and are therefore unlikely to possess consciousness. Meanwhile, Bernard Baars and Stanislas Dehaene's Global Workspace Theory (GWT) holds that consciousness arises from the global broadcasting of information in the brain's "workspace" — under this framework, if AI systems implement a similar global information integration architecture, the emergence of consciousness is not theoretically impossible. But these theories themselves remain under intense debate and have not been decisively verified by experiments.

When large models generate fluent text, we easily anthropomorphize them, thinking they "understand" the content. But from an engineering perspective, they are merely predicting the probability distribution of the next token. This "Chinese Room"-style challenge remains unresolved: Can symbol manipulation produce genuine understanding?

The Chinese Room is a famous thought experiment proposed by philosopher John Searle in 1980: Imagine a person who doesn't understand Chinese sitting in a room, processing Chinese symbol inputs and producing reasonable Chinese answers by following a comprehensive rule book. From the outside, the room appears to "understand" Chinese, but the person inside actually knows nothing about Chinese. Searle used this to argue that pure symbol manipulation (at the syntactic level) cannot produce genuine semantic understanding, thereby challenging the position of Strong AI — that an appropriately programmed computer can truly possess a mind. Closely related is Stevan Harnad's "Symbol Grounding Problem" proposed in 1990: How do symbols in a symbol system acquire correspondence with the real world? Current multimodal large models, by simultaneously processing text, images, audio, and other sensory data, are seen by some researchers as a step toward symbol grounding, but whether this constitutes genuine "grounding" — truly connecting symbols to meaning in the world — remains widely debated.

The Realistic Boundaries of Computer Scientists Building a Brain

Returning to the original question, the answer depends on how we define "brain":

• If referring to a functional intelligence system: Computer scientists are already on the way, with remarkable achievements. Modern AI's performance on specific tasks has already demonstrated certain "brain-like" capabilities.
• If referring to a complete biological brain simulation: This requires deep collaboration among computer science, neuroscience, and biology, far beyond what computer scientists can accomplish alone.
• If referring to a conscious entity: This may exceed the explanatory power of current scientific paradigms, involving unresolved philosophical puzzles.

Building a "brain" has never been a purely computer science problem. It requires neuroscientists to provide biological insights, cognitive scientists to understand thinking mechanisms, and philosophers to clarify conceptual boundaries. True breakthroughs often occur at the intersections of disciplines.

Conclusion: From Replicating Brains to Creating New Forms of Intelligence

This question remains perennially relevant because it simultaneously inspires human ambition and humility. We yearn to create intelligence on par with our own, yet must admit that our understanding of our own brains remains shallow.

Perhaps a more constructive approach is this: computer scientists need not "replicate" the brain, but can instead create an entirely new form of intelligence — one that doesn't necessarily work like the human brain, but can equally think, create, and solve problems. Airplanes don't need to flap wings to fly, and artificial intelligence may not need neurons to be intelligent.

This classic flight analogy originates from early AI research discourse, often used to illustrate that functional equivalence doesn't require mechanistic identity. Early aviation pioneers did attempt to imitate birds' flapping flight (such as da Vinci's ornithopter designs), but modern aircraft use fixed wings and jet propulsion to achieve flight performance far exceeding that of birds. Similarly, the mainstream approach in current AI research doesn't pursue simulating the precise behavior of biological neurons, but instead seeks alternative computational solutions for achieving intelligent functions. However, it's worth noting that this analogy has its deep limitations: the physical principles of flight (aerodynamics) are fully understood, while the nature of intelligence remains an open question. Can we be confident that intelligence, like flight, can be achieved through entirely different mechanisms? This itself is a question that needs answering, not a premise that can be assumed.

This possibility of "heterogeneous intelligence" is perhaps the most exciting direction. It reminds us that the ultimate goal of building a brain may not be to replicate nature, but to expand the very definition of intelligence itself.