I got curious about a number this week. If the human brain has roughly 100 trillion synapses, and each synapse is loosely analogous to a parameter in an artificial neural network, what does that tell us about where AI actually stands?
The short answer: further away from the human brain than most headlines suggest, and closer than most people realise.
The numbers
The human brain contains approximately 86 billion neurons. Each neuron forms somewhere between 1,000 and 10,000 connections to its neighbours, giving a total of roughly 100 to 500 trillion synapses. The most commonly cited figure is 100 trillion, though some neuroscientists put it higher. A quadrillion is not out of the question.
In an artificial neural network, the equivalent of a synapse is a parameter, a numerical weight that determines how strongly one artificial neuron influences the next. So on a crude, back-of-the-envelope basis, the human brain operates with something like 100 trillion parameters.
The largest AI models we know about are not close to that number. GPT-4’s architecture has never been officially confirmed by OpenAI, but credible estimates put it at roughly 1.8 trillion parameters across its mixture-of-experts system. Meta’s Llama 4 Behemoth, previewed in April 2025, is reportedly a 2 trillion parameter model. DeepSeek-R1, the open-weight reasoning model released in January 2025, has 671 billion parameters.
So the biggest current models sit at roughly 1 to 2 trillion parameters. The brain sits at 100 trillion or more. That is a gap of roughly 50 to 100 times, or about two orders of magnitude.
Why the comparison is useful but imperfect
The synapse-to-parameter analogy is the best shorthand we have, but it breaks down under scrutiny. A biological synapse is not a single number. It is a complex molecular assembly containing around 2,000 variable proteins, capable of fast and slow learning, modulated by neurotransmitters, and responsive to context in ways that a floating-point weight cannot replicate. As Princeton researchers have cautioned, “we should not commit to an argument for scarcity of computational resources as long as we poorly understand the computational machinery in question.”
There is also the question of architecture. The brain is not one giant neural network. Beren Millidge, a researcher who has done some of the most careful quantitative work on brain-to-AI comparisons, describes it as “a set of about 10 to 15 semi-independent specialised models, each ranging from about the size of a large language model up to maybe an order of magnitude larger, with extremely high bandwidth interconnect.” The visual cortex alone, which occupies roughly 30% of the cortical volume, would contain an estimated 3 to 5 trillion synapses. The language regions, Broca’s and Wernicke’s areas, take up only about 6% of the cortex, putting the brain’s dedicated language-processing capacity at somewhere around 400 to 700 billion parameters.
That last number is striking. It sits right in the range of current large language models. If you accept the crude parameter equivalence, it suggests that today’s frontier AI models are at rough parity with the brain’s language centres, but are orders of magnitude smaller than the brain as a whole. The gap is not in language. It is in everything else: vision, spatial reasoning, motor control, memory, social cognition, the full stack of general intelligence.
The energy gap is even more dramatic
Here is where the comparison gets properly humbling. The human brain runs on about 20 watts, roughly the power draw of a dim light bulb. It consumes about 25% of the body’s total energy budget, which tells you something about the evolutionary cost of intelligence, but it still runs on a banana and a sandwich.
Training GPT-4, by contrast, consumed an estimated 50 gigawatt-hours of energy, enough to power San Francisco for three days, and cost over 100 million dollars. Current-generation training runs for models comparable to GPT-4o draw around 20 to 25 megawatts of power each, sustained over roughly three months. That is enough to power 20,000 American homes.
The gap in energy efficiency between biological and artificial intelligence is estimated at somewhere between 10,000 and 225,000 times, depending on how you measure it. The brain achieves general intelligence on 20 watts. AI achieves narrow excellence on megawatts. Geoffrey Hinton, the Nobel Prize-winning pioneer of deep learning, has said that digital systems may simply be “a better form of intelligence than people.” But even he would acknowledge they are a spectacularly less efficient one.
What the AI pioneers think
The question of whether AI will reach brain-scale, and whether brain-scale even matters, is dividing the field’s most prominent figures.
Sir Demis Hassabis, CEO of Google DeepMind, argues that brains are “the most exquisite and complex phenomena we know of in the universe” and that they are “in fact extremely general.” He views the human brain as an approximate biological Turing machine, and believes modern AI models are increasingly approaching that level of generality.
Yann LeCun, Meta’s outgoing chief AI scientist and a co-winner of the 2018 Turing Award, takes the opposite view. He has called the concept of general intelligence “complete BS,” arguing that human intelligence is highly specialised and shaped by biology to deal with specifically human challenges. “We think of ourselves as being general,” LeCun has said, “but it is simply an illusion.”
Hinton, who shared the 2024 Nobel Prize in Physics for his foundational work on neural networks, sits somewhere between these positions but has become increasingly concerned about what happens when AI systems do reach brain-scale capability. “If you want to know what it is like not to be the apex intelligence,” he said in a 2025 interview, “ask a chicken.”
What this means for the next five years
Current trajectory suggests we will see models in the 10 to 30 trillion parameter range within the next few years. That would put them at roughly 10 to 30% of the brain’s estimated parameter count, though the comparison will grow less meaningful as architectures diverge further from simple parameter counting.
The more interesting metric may not be parameter count at all but capability per watt. The brain’s extraordinary energy efficiency is driving a parallel research effort in neuromorphic computing, where chips are designed to mimic biological neural architectures. Intel’s Loihi 2 chip is reportedly 10 times more energy-efficient than equivalent GPU processing; IBM’s NorthPole chip, which eliminates the traditional separation between memory and computation, is 25 times more efficient for image tasks.
The race, in other words, is not just to build bigger models. It is to build smarter ones. And the benchmark, whether we like it or not, remains the three-pound organ sitting between your ears that runs on roughly the same power as a kitchen night light.
Research notes
Beren Millidge, “The Scale of the Brain vs Machine Learning,” August 2022. https://www.beren.io/2022-08-06-The-scale-of-the-brain-vs-machine-learning/
UCLA Brain Research Institute, “Billions of neurons, trillions of synapses.” https://bri.ucla.edu/brain-fact/billions-of-neurons-trillions-of-synapses/
Epoch AI, “Over 30 AI models have been trained at the scale of GPT-4,” June 2025. https://epoch.ai/data-insights/models-over-1e25-flop
Epoch AI, “How much energy does ChatGPT use?” https://epoch.ai/gradient-updates/how-much-energy-does-chatgpt-use
MIT Technology Review, “We did the math on AI’s energy footprint,” May 2025. https://www.technologyreview.com/2025/05/20/1116327/ai-energy-usage-climate-footprint-big-tech/
Geoffrey Hinton, Nobel Prize podcast, 2024. https://www.nobelprize.org/prizes/physics/2024/hinton/podcast/
Sir Demis Hassabis on X, June 2026. https://x.com/demishassabis/status/2003097405026193809
University of Sydney, “How the human brain is inspiring energy-efficient AI,” August 2024. https://www.sydney.edu.au/news-opinion/news/2024/08/16/how-the-human-brain-is-inspiring-energy-efficient-ai.html
ScienceDaily, “Artificial intelligence uses less energy by mimicking the human brain,” March 2025. https://www.sciencedaily.com/releases/2025/03/250326123554.htm
Meta AI, “The Llama 4 herd,” April 2025. https://ai.meta.com/blog/llama-4-multimodal-intelligence/
LessWrong, “Parameter vs Synapse?” https://www.lesswrong.com/posts/7htxRA4TkHERiuPYK/parameter-vs-synapse