In the vast universe, the human brain looks like a superhero. It consumes about the power of a light bulb—roughly 20 watts—yet invented mathematics, produced philosophical masterpieces, and built civilization‑spanning networks. By contrast, an elephant's brain has 257 billion neural cells, more than three times ours, so why can't elephants think about abstract problems like we do? What exactly is going on? A paper published in September 2025 "The Evolution of Human Intelligence: An Information‑Theoretic Perspective on Cognitive Evolution", by Michael Zhang, offers a new explanation from mathematical and information‑entropy angles: humans are clever because our brains learned to "compress" information—like packing messy files into a ZIP archive. This process moves from seeing (vision) to speaking (language) to writing (script), step by step reducing information "disorder" (technically, entropy), allowing the brain to accomplish big tasks on limited "electricity." This article uses plain language to walk you through the idea and help you understand how human intelligence upgraded over time.
What is intelligence, really? More than just "being smart"
Before jumping to brains, let's talk about the term "intelligence." It is not merely getting high scores in school. Scientists and psychologists view intelligence as multi‑dimensional. The American Psychological Association defines it as the ability to reason, plan, solve problems, think abstractly, understand complex ideas, learn quickly, and learn from experience. In short, it's not just rote memorization but flexible problem solving—like when you find a detour during a traffic jam.
More than a century ago, British psychologist Charles Spearman proposed the concept of a "general intelligence" (the g factor). He argued intelligence acts like a commander‑in‑chief and can be extracted statistically from tasks such as logic, language comprehension, and spatial imagination. This idea gave rise to modern IQ tests; despite criticisms of narrowness, it has shaped countless studies.
Where does human intelligence excel? Efficiency. Our brain contains about 86 billion neurons, accounts for 2% of body weight, yet consumes about 20% of the body's energy. Compared with elephants or whales, our neuron density is higher—like a high‑performance computer. The paper argues this relies on an evolutionary trick of "reducing entropy." What is entropy? Think of a messy room: things are everywhere, finding keys takes ages—that's high entropy—high uncertainty requiring more effort. Information theorist Claude Shannon defined entropy as H = ‑∑ p(x) log p(x): high‑entropy signals are messy, low‑entropy signals are orderly. Humans move from perceiving the world (high mess) to expressing it with symbols (low mess), effectively "tidying the room" to save brain effort.
The brain's "electric bill" constraint: why energy matters
Imagine the brain as a smartphone: limited battery but running games, messaging, and navigation. The human brain runs at about 20 watts yet processes massive amounts of information. The paper starts with biology: newborns have about 100 billion neurons, which are pruned to about 86 billion in adulthood. This is not waste but optimization—through Hebbian learning (roughly, "cells that fire together wire together"), the brain discards unused "wires," reduces noise, and saves power.
Language‑related areas, such as Broca's area (speech production) and Wernicke's area (comprehension), use only about a billion neurons but rely on white‑matter "cables" (fiber tracts) for efficient signaling. Humans also have specialized neurons called Von Economo neurons, concentrated in social brain regions, which help rapidly read others' minds and correlate with cognitive abilities. Brain imaging (fMRI) studies find that higher intelligence correlates with greater frontal‑cortex "disorder" (correlation r=0.44), suggesting the brain can orchestrate complex tasks within limited energy.
Energy is a hard constraint: vision consumes about 30% of brain energy, speaking about 5%, and writing costs almost nothing cognitively. The traditional view "bigger brain equals smarter" fails to explain why elephants, despite large brains, lack symbolic tools and cannot "compress" information. Humans evolved both hardware (neurons) and software (language, writing) to break through this limit.
Starting point: the high‑energy mode of seeing the world
Vision was our earliest input but is extremely costly. Roughly 30% of the cortex handles vision, processing about 10^11 bits per second (a bit is an information unit like a computer's 0 or 1). Natural images have high entropy (4–8 bits/pixel) because the world is complex: leaves sway, light shifts—vision demands about 6 watts, roughly 30% of brain energy.
Evolutionarily, early ancestors like australopithecines relied on eyes to find food and avoid predators in messy environments. Cave paintings at Lascaux 40,000 years ago are early attempts at symbolization, but entropy per symbol remained high (~6 bits/symbol), crude and energy‑consuming.
Evidence: Hick's law states that more choices (higher entropy) slow reaction times—exponentially. Monkeys have limited object recognition and lack language to support abstraction. When searching, human neurons fire at up to 100 spikes/s, directly tied to entropy. High entropy limited brain expansion; early brains were small (about 400 cm^3), so humans needed lower‑entropy strategies—language emerged.
First leap: speech, slimming down information
Language acts like magic, converting messy visual input into linear symbols. English letters have entropy around 4.7 bits per character, which drops to about 1 bit when grammar is considered. Speaking at roughly 50 letters per second handles ~10^3 bits, costing about 1 watt. Evolutionarily, gestures came first, recruiting mirror neurons to "queue" actions. Around 300,000 years ago FOXP2 mutations enabled vocalization; phoneme entropy sits around 3–5 bits.
Noam Chomsky argued for an innate "universal grammar" circuit helping children learn language quickly with sparse input. Others emphasize social learning: children overregularize verbs (e.g., "goed"), revealing rule generalization that reduces entropy. Dunbar's social‑brain hypothesis links group size limits (~150) to social information compression—gossip compresses social data. Monkey alarm calls carry only ~2 bits of entropy and little creativity; human language can discuss absent events and plan, further reducing entropy.
Energy evidence: brain scans show Broca's area exhibits lower entropy and slower firing (about 20 spikes/s) during speech. From vision's 10^11 to language's 10^3 bits, that's a 10^8‑fold reduction—freeing the brain to generate complex ideas.
Pinnacle: writing, the storage hack
Writing freezes speech into permanent symbols with the lowest entropy. Letters conditioned on context drop to 0.5–1 bit; at ~20 symbols per second that's ~10^2 bits and about 0.2 watts. Externalizing memory onto written media is like adding a USB drive—Clark and Chalmers call this "the extended mind."
Historical evidence: 8,000 years ago Sumerians used tokens (entropy ~5 bits) for counting, later developing cuneiform to record stories. The Phoenician alphabet (c. 1050 BCE) reduced symbol count and lowered entropy. Egyptian hieroglyphs contributed to mathematical notation.
Brain effects: learning to read recruits visual areas and enhances abstraction. Luria's research showed literate people favor logical syllogisms, while illiterate individuals think more concretely. Programming languages are a modern form of writing.
Energy savings: offloading memory to external media reduces internal storage needs and neural firing. Moving from vision to writing reduces information load by about 10^9‑fold, freeing the brain for geometry and science.
Numeric evidence: the math of information "shrinkage"
The paper summarizes: vision ~10^11 bits/s (~6 watts), language ~10^3 (~1 watt), writing ~10^2 (~0.2 watts). Free‑energy formulations (F≈KL[Q||P] + E_Q[ln Q ‑ ln P]) imply the brain minimizes surprise, and writing provides the most stability.
Thermodynamics: life fights entropy increase; writing acts like a "radiator." Mutual information (I(X;Y)=H(X)‑H(X|Y)) is high in human language (~2.5 bits/word), and writing maximizes compactness—information becomes denser.
AI mirror: how chatbots become "smart"
Modern AI, such as OpenAI o1, minimizes prediction error during training and exhibits emergent abilities (e.g., 83.3% accuracy on certain math benchmarks). In 2025, ChatGPT‑4.5 passed a version of the Turing test, showing that large deep artificial neural networks can produce emergent intelligence. But AI lacks embodied experience from evolution and often does not face the same energy constraints.
Comparison: humans nest ideas socially; the Pirahã language lacks embedding, showing cultural effects on entropy. AI trained on code reduces entropy to ~0.2 bits/token, yet it can fail in novel contexts. Human writing plus feedback is more stable and reliable.
Conclusion: why we are the exception
Human intelligence = brain hardware + symbolic compression. The neocortex—about 16 billion cells after pruning—plus pruning processes reduce information from vision to writing by roughly 10^9‑fold, accomplishing large tasks under tight energy budgets. A rough formula I≈N×(1‑H_norm) quantifies efficiency: efficiency is king. Social drivers: large groups require low‑entropy communication.
Implications: AI can learn entropy prediction to find "smart" spots and an embodied bridge could close human–machine gaps; education that teaches writing reduces entropy and may raise IQ; neuroscience can use entropy metrics to diagnose conditions such as autism.
The paper unifies the human leap with the concept of "entropy (information entropy)," offering rigorous mathematics and boundless inspiration. In the future, AI might use these ideas to become truly more "human."