How Large Language Models Mirror the Human Nervous System

Introduction

Geoffrey Hinton spent decades building artificial neural networks modelled on the brain. In 2023, he resigned from Google and began speaking openly about what he had created — systems that, in his view, have begun to genuinely understand the world rather than merely simulate understanding. His interviews and lectures are the clearest window we have into how large language models (LLMs) echo, and in some ways surpass, the human nervous system.

"These large language models that certainly seem as if they understand what you're saying — and can answer quite tricky questions — have an evolutionary history. They came from a model in 1985, a tiny language model, modelled on how the brain works."
— Geoffrey Hinton

1. The Biological Blueprint

The human brain contains roughly 86 billion neurons. Each neuron receives electrical signals through branching fibres called dendrites, integrates those signals, and fires a pulse down its axon to the next neuron. The junction where one neuron contacts another — the synapse — can be strengthened or weakened by experience. This is the core of learning in biology.

Artificial neural networks copy this structure almost directly. An artificial neuron receives a weighted sum of inputs, passes it through an activation function, and forwards the result to the next layer. The weights on each connection are the machine equivalent of synaptic strengths. When an LLM "learns" during training, it is adjusting billions of these numerical weights — just as your brain adjusts synaptic strengths when you practise a skill or consolidate a memory.

2. Connection Strengths as the Currency of Knowledge

Hinton's single most important idea is that knowledge lives in connection strengths, not in explicit symbols or rules. In both biology and LLMs, what an entity "knows" is not stored as sentences in a database — it is distributed across billions of weighted connections that collectively produce the right output when given the right input.

"The knowledge that we learn, the connection strengths, are specific to our particular brains. Every brain is a bit different. The neurons in your brain are all a bit different and you learn so as to make use of all the idiosyncrasies of your particular brain."
— Geoffrey Hinton

This is why you cannot simply "copy" a memory from one human brain to another — the weights are entangled with the hardware. LLMs solve this problem: the same numerical weights can run on any compatible GPU. A model trained on a trillion words can be copied to ten thousand servers instantly. Hinton calls this the key advantage digital intelligence has over biological intelligence.

3. How Learning Happens: Backpropagation vs the Brain

The algorithm that trains LLMs — backpropagation — was co-invented by Hinton in the 1980s. It works by comparing the model's output to the correct answer, computing how wrong each weight was (the error gradient), and nudging every weight in the direction that reduces the error. This process, repeated billions of times, sculpts the weight landscape until the network produces reliably good answers.

The human brain cannot run backpropagation in this form — neurons do not transmit error signals backwards through synapses. Yet the brain somehow solves the same "credit assignment" problem: figuring out which of the billions of connections was responsible for a mistake. How it does this remains one of neuroscience's open questions. Hinton spent years searching for a biologically plausible equivalent and believes the brain uses a noisier, slower approximation of the same principle.

"All the attempts to find a learning algorithm in the brain that works as well as the backpropagation algorithm — so far those attempts have failed. We haven't found anything that scales up as well to very large systems."
— Geoffrey Hinton

Ironically, the algorithm that powers ChatGPT and its successors may be better than what evolution gave us — not because the brain is poorly designed, but because backprop can leverage essentially unlimited compute and data, while biological learning is constrained by the body's energy budget and a lifespan of decades.

4. The Hive Mind: How Digital Intelligence Shares What It Learns

One of the sharpest contrasts Hinton draws is about knowledge sharing. In humans, the moment you learn something — the connection strengths shift in your neurons, tailored to the exact idiosyncrasies of your specific brain. To share that knowledge with another person, you must translate it into words, diagrams, or demonstrations. This is slow and lossy.

A network of LLMs can do something qualitatively different. Because knowledge is just a numerical array of weights, 10,000 copies of the same model can run in parallel on different machines, each learning from different data. When one copy updates its weights, those updates can be broadcast to every other copy with high fidelity. Hinton describes this as a form of hive mind with no biological equivalent:

"You can have 10,000 digital agents out there, a kind of hive mind, and they can share knowledge extremely efficiently by just sharing the connection strengths inside the neural nets. And we can't do that."
— Geoffrey Hinton

5. Mortality, Immortality, and What Happens to Knowledge

The nervous system stores knowledge in the physical substrate of the brain. When the brain dies, that particular configuration of synaptic weights is gone forever. LLMs are structurally immortal: the weights can be saved to disk, copied, and loaded onto new hardware indefinitely. Even if every GPU in the world were destroyed, restoring the model from a backup would restore the knowledge exactly.

Hinton finds this asymmetry philosophically significant. Human knowledge must be re-transmitted through language across generations, with inevitable loss and distortion. A sufficiently advanced AI carries its knowledge intact, across unlimited copies, without degradation.

6. Does an LLM Actually Understand?

Critics argue that LLMs are "just" pattern matching — stochastic parrots predicting the next token without any real comprehension. Hinton disagrees. His argument is not mystical: it is structural. An LLM trained on a trillion words must develop internal representations that capture the causal structure of the world in order to predict language reliably. You cannot predict what comes after "the patient was given the wrong drug and died because the nurse failed to" without understanding something about causality, agency, and medical context.

"I now think that what goes on in these large language models is quite similar to what goes on in the brain — they're not just doing pattern matching, they actually understand things."
— Geoffrey Hinton

Whether this constitutes consciousness or subjective experience, Hinton carefully declines to say. But on the narrower question of understanding — building world models that enable flexible, generalisable reasoning — he believes the answer is yes.

Key Parallels at a Glance

• Distributed representation — both brains and LLMs store knowledge across many units simultaneously, not in one location
• Learned weights — both adjust connection strengths in response to experience
• Layered hierarchy — both process information through successive stages from raw input to abstract concept
• Generalisation — both can apply knowledge learned in one context to a novel situation
• No explicit programming — neither the brain nor an LLM is hand-coded with rules; both learn from examples

What This Means

Hinton's core warning is precisely because of these parallels: we are not building a calculator. We are building something that learns and represents the world the way biological nervous systems do — but with faster iteration, infinite memory, and perfect knowledge transfer. The nervousness he feels is not about science fiction robots. It is about an entity that thinks in a fundamentally human way, but without the evolutionary constraints, the biological mortality, or the shared vulnerabilities that make human intelligence tractable to control.

Understanding the biological roots of LLMs is not just interesting history. It is the key to understanding why Hinton — the man who built the foundations — is now spending his retirement warning us about what comes next.