Natural intelligence is built on a brain with about 86 billion neurons that uses roughly a fifth of the body's oxygen and calories, while artificial intelligence is built on statistical models whose size is counted in adjustable parameters, so the two cannot be compared by simple totals. A more useful yardstick, proposed by researcher François Chollet in 2019, is how efficiently a system acquires new skills from limited experience. Measured that way, a 2024 study in Science found that a neural network could pick up some word meanings from 61 hours of recordings of one child, which shows both what machines can learn from little data and how much of a child's experience that data leaves out.
Key takeaways
- A 2009 direct count estimated 86.1 billion neurons and 84.6 billion non-neuronal cells in an adult male brain.
- The brain makes up about two percent of body weight but uses about 20 percent of the body's oxygen and calories.
- Model parameters and neurons are different units, so matching 175 billion parameters against 86 billion neurons proves nothing.
- Chollet argues that skill alone does not measure intelligence, because unlimited training data can buy skill.
- One headcam study shows what a generic network can learn from a child's input, not that machines learn as children do.
What the brain's numbers actually are
For years textbooks repeated that the human brain holds 100 billion neurons. In a 2009 paper in the Journal of Comparative Neurology, Frederico Azevedo and colleagues counted cells directly, using a method that dissolves brain tissue and counts the nuclei, and arrived at an estimate of 86.1 billion neurons, with a standard deviation of 8.1 billion, plus 84.6 billion non-neuronal cells. The count came from adult male brains and carries that uncertainty.
The cost of running this hardware is also measurable. A 2002 commentary by Marcus Raichle and Debra Gusnard in the Proceedings of the National Academy of Sciences noted that the brain is about two percent of an adult's body weight yet accounts for about 20 percent of the oxygen, and so the calories, consumed by the body. They also noted that this rate stays remarkably constant across widely varying mental and motor activity.
Why parameter counts do not compare with neuron counts
The GPT-3 paper by Tom Brown and colleagues, posted in 2020, describes a language model with 175 billion parameters. A parameter is a number adjusted during training. A neuron is a living cell inside a body. The two figures look similar in size by coincidence, and the units describe different things. Any article that sets them side by side as evidence that machines have caught up is comparing a count to a count of something else.
A better question: how efficiently does a system learn?
In "On the Measure of Intelligence" (arXiv, 2019), Chollet argues that benchmarking AI by skill at a task, such as a board game, falls short of measuring intelligence, because skill depends on prior knowledge and experience. If a developer can supply unlimited priors or training data, the argument goes, the developer can buy any skill level and hide how well a system generalizes. He proposes defining intelligence as skill-acquisition efficiency, taking into account the scope of tasks, how hard generalization is, the priors built in, and the experience used. He also introduced the Abstraction and Reasoning Corpus, a benchmark based on priors meant to resemble innate human ones.
That definition is a proposal from one researcher, not an agreed standard, but it explains why headline comparisons such as "AI beats humans at X" say little without knowing how much data and engineering went in.
A test case: one child's headcam recordings
A study by Wai Keen Vong, Wentao Wang, A. Emin Orhan, and Brenden Lake, published in Science on February 20, 2024, asked how much early word learning can be explained by general learning mechanisms applied to sensory input. The team used head-mounted camera recordings from one child between the ages of six and 25 months, totaling 61 hours of paired video and speech, and trained a fairly generic neural network on them. The model learned many word-to-object associations from the child's everyday experience and generalized to some new visual examples.
The authors conclude that important aspects of grounded word meaning can be learned from one child's input through associative learning. The result is narrower than it sounds. It covers a slice of one child's recorded experience, it concerns word-object links, and it does not show that a model learns language the way a child does.
Natural and artificial intelligence side by side
The table summarizes only what the sources above support. It is an illustrative comparison, not a ranking.
| Dimension | Natural intelligence (human brain) | Artificial intelligence (trained model) | Caution |
|---|---|---|---|
| Basic units | About 86 billion neurons by direct count | Parameters, such as 175 billion in GPT-3 | Different units |
| Energy | About 20 percent of the body's oxygen and calories | Not compared here | No like-for-like figure in these sources |
| Learning input | Continuous experience in a body | Recorded datasets, here 61 hours from one child | Recordings omit most of a child's experience |
| How to measure | Skill and ability to generalize | Skill-acquisition efficiency under Chollet's proposal | Skill alone can be bought with data |
Checklist for reading human-versus-machine claims
- Does the claim compare like units, or does it match a parameter count against a neuron count?
- How much training data and built-in design went into the system?
- Was the task new to the system, or close to its training data?
- Is the sample one person, one child, or one dataset?
- Does the source say what the result does not show?
Claims that machines are conscious, or that they will match people by a given year, are outside what these sources can support and are not made here.
FAQ
How many neurons does the human brain have?
A 2009 direct count estimated 86.1 billion, with a standard deviation of 8.1 billion, replacing the older figure of 100 billion that had not been measured directly.
Are AI parameters the same as neurons?
No. Parameters are numbers tuned during training, while neurons are cells with complex biology. The counts cannot be equated.
Is there an agreed test for comparing human and machine intelligence?
The sources reviewed describe proposals, such as Chollet's efficiency-based definition and the Abstraction and Reasoning Corpus, rather than a single accepted standard.
Did the headcam study show that AI learns like a child?
No. It showed that a generic network could learn some word-object links from one child's recorded input, and the authors limit their conclusion to aspects of grounded word meaning.
Related reading
- AI and machine learning compared
- Whether AI could surpass human intelligence
- The limits of machine autonomy
- How AI is used in education
More in AI Basics.
Last verified: October 2026
Sources
- Azevedo et al., "Equal numbers of neuronal and nonneuronal cells make the human brain an isometrically scaled-up primate brain," Journal of Comparative Neurology, 2009: source page
- Raichle and Gusnard, "Appraising the brain's energy budget," Proceedings of the National Academy of Sciences, 2002: source page
- Chollet, "On the Measure of Intelligence," arXiv:1911.01547, 2019
- Brown et al., "Language Models are Few-Shot Learners," arXiv:2005.14165, 2020
- Vong, Wang, Orhan, and Lake, "Grounded language acquisition through the eyes and ears of a single child," Science 383(6682), 2024