In 2024, AI won two Nobel prizes in a single year. Physics went to Geoffrey Hinton and John Hopfield for neural networks; Chemistry went to Demis Hassabis and John Jumper for AlphaFold. Neural networks were inspired by the brain. So how close are they to the real thing? Closer in some ways than you’d think, and astonishingly far in others.

Read this first. Brain figures are scientific estimates with wide ranges, and comparing a synapse to a model parameter is an analogy, not an equivalence. Numbers marked [Estimate] are my own rough calculations.

The warning from the man who built it

Hinton left Google in 2023 so he could speak freely about AI risk, and the Nobel gave him a bigger stage. His concerns fall into four groups: disinformation that makes truth hard to tell apart from fakes; job losses in routine knowledge work, with the gains concentrated in few hands; misuse, from cyberattacks to autonomous weapons; and, most seriously, losing control of systems smarter than us. He has put the chance of an AI-driven catastrophe in the coming decades at roughly 10% to 20%. Not everyone agrees: Yann LeCun, another pioneer, considers that fear overblown.

One reason the debate is hard: we still don’t fully understand what happens inside either kind of mind.


Link to Dashboard

Connections: brain vs open models

To keep the comparison fair, I use open-weight models, whose sizes are public, rather than closed frontier models, whose sizes are not disclosed.

  • Human brain: about 86 billion neurons and roughly 100 to 1,000 trillion synapses.
  • DeepSeek V4-Pro: 1.6 trillion parameters in total, about 49 billion active per token, using a Mixture-of-Experts (MoE) design. It needs an eight-GPU server.
  • Zhipu GLM-5.2: 744 billion parameters in total, about 40 billion active per token.

Even the largest open model has 60 to 600 times fewer “connections” than a human brain. And a synapse is far richer than a parameter: it changes over time, uses chemistry as well as electricity, and works with timing, not just weights.

Energy: the gap that matters most

  • Brain: about 20 watts, less than a dim light bulb.
  • Eight-GPU server running a model like DeepSeek V4-Pro: around 10 kilowatts. [Estimate] That is roughly 500 times the brain.
  • El Capitan, the fastest supercomputer: about 30 megawatts at peak.

Common estimates put the brain somewhere between 1015 and 1018 operations per second. At the high end, it matches an exascale supercomputer on about a millionth of the power.

Speed: where silicon wins

A neuron fires at most a few hundred times per second. A GPU performs billions of operations per second per core. In raw arithmetic, the computer wins by many orders of magnitude, and it can read every book ever scanned. The brain compensates with massive parallelism and by doing far less work per decision.

A surprising similarity: sparsity

Modern MoE models activate only a small slice of their parameters for each token, about 3% in DeepSeek V4-Pro. The brain works the same way: only a small fraction of neurons are active at any moment. Both save energy by not using everything all the time. That’s convergent design, not coincidence.

Where the brain is still unbeaten

  • Learning from few examples. A child learns “dog” from a handful of dogs. A model needs vast amounts of data, and its training run consumes energy measured in gigawatt-hours.
  • Learning all the time. The brain updates itself continuously. A model is mostly frozen after training.
  • Memory and compute together. In the brain, the synapse stores and computes. In a GPU, a large share of the energy goes into moving data between memory and the chip.

Auditing a mind

We can’t fully read either one. But AI has one big advantage over neuroscience: researchers can record every artificial neuron exactly, pause a model mid-thought, copy it, and change one value to see what happens. No neuroscientist can do that to a living brain. Even so, interpretability research explains only a small part of what large models do inside. That is a real problem when some reasoning models produce shorter, less readable chains of thought.

Closing the gap: brain-inspired chips

Neuromorphic computing tries to copy the brain’s design: spiking neurons, sparse activity and memory placed next to compute. Intel’s Hala Point, the largest such system, packs 1,152 Loihi 2 chips with 1.15 billion artificial neurons and 128 billion synapses into a six-rack-unit chassis, drawing at most 2,600 watts. That’s about 1% of a brain’s neurons. These chips shine on sparse, event-driven tasks, but they don’t run today’s large language models yet.

The takeaway

Today’s AI beats the brain on speed and on breadth of knowledge. The brain wins, by a huge margin, on energy, data efficiency and continuous learning. The next big leaps in AI may come less from making models bigger and more from making them work a little more like the 1.4 kg of tissue that runs on 20 watts.

Sources

Written November 2026. Brain figures are estimates and vary between sources.

Tags: , , , , ,

Leave a Reply

Your email address will not be published. Required fields are marked *