If we compare the human brain to a computer, its energy efficiency is astonishing. Our biological hardware—the brain—consumes only around 20 watts, roughly the power of a dim incandescent bulb. Yet this modest energy budget supports mathematics, philosophy, and even the construction of global civilization.
By contrast, nature includes brains with seemingly more “luxurious” hardware. For example, the elephant brain contains as many as 257 billion neurons, about three times the human total (86 billion). Why, then, has the elephant not developed complex mathematics or written language?
Two recent scientific studies—one on the evolutionary entropy reduction underlying human cognition and another comparing the brain with large language models (LLMs)—help unlock the answer: the essence of human intelligence lies not in raw hardware, but in extreme information “compression” and “prediction.”
1. Evolution’s accounting: how to “cheat” with 20 watts
To understand why humans are smart, we must first audit the brain’s energy ledger like an accountant. According to research by Michael Zhang (2025), the evolution of human intelligence is essentially a history of maximizing semantic energy efficiency.
Under the strict 20‑watt budget, the brain must economize:
1. The high cost of vision: The most primitive mode of cognition is vision. Processing visual information consumes a great deal of energy because natural images are full of redundancy and noise (high entropy). To handle raw data that can reach up to 100 billion bits per second, the visual cortex requires roughly 6 watts, with low efficiency—about 16.7 bits per joule. If humans relied solely on perception for thought, the brain would have “crashed.”
2. Language as dimensionality reduction: The emergence of language was the first leap. By compressing complex visual scenes into linear sound symbols, humans reduced processing energy to around 1 watt.
3. Writing as the ultimate compression: The invention of writing was a true “cheat code.” It fixed ephemeral speech into stable symbols and outsourced memory and computation to paper or stone. Studies estimate that the semantic processing energy of writing systems requires only about 0.2 watts, boosting efficiency to some 500 bits per joule.
By reducing entropy from vision to language and then to writing, humans freed the brain from heavy real‑time data processing and achieved the highest intelligence with minimal energy. Elephants, despite their many neurons, lack such symbolic compression mechanisms and remain trapped in high‑energy sensory processing.
2. The surprising resonance between brain and AI: hierarchy and time in duet
While evolution explains why we process information this way, neuroscience reveals how the brain operates. Remarkably, the human brain’s processing pattern closely mirrors that of prominent large language models (LLMs such as GPT‑2 and Llama‑2).
A study by Goldstein et al. (2025) published in Nature Communications used high‑precision intracranial electroencephalography (ECoG) to record human brain activity while subjects listened to stories and compared those signals with the internal workings of AI models.
The researchers found a striking correspondence: an AI model’s “depth” maps onto the brain’s “time.”
In models like GPT, information is processed through successive neural layers. Shallow layers handle simple forms; deeper layers capture complex semantics and context. The team observed that when humans hear a word, the brain’s temporal response—especially in language‑related areas like Broca’s region—recapitulates the model’s layered structure:
• AI shallow layers (layers 1–10): correspond to the brain’s very short‑latency neural responses after hearing a word.
• AI deep layers (layers 30–48): correspond to later neural activity (hundreds of milliseconds after hearing the word).
This implies the brain does not instantaneously grasp a word’s full meaning; rather, it unfolds information over time, layer by layer, much like an AI peels an onion. This sequential, progressively non‑linear transformation demonstrates a fundamental computational similarity between current LLMs and the human brain.
3. The prediction engine: minimizing “surprise”
Both studies converge on the importance of prediction.
From an information‑theoretic perspective, the brain is a machine dedicated to “minimizing free energy.” The external world is full of uncertainty (high entropy), which poses great survival pressure on a 20‑watt brain. To survive, the brain must continuously predict what will happen next to reduce “surprise.”
• Writing removes uncertainty through fixed symbols, acting as an “entropy anchor” so the brain need not guess at every turn.
• Neural networks, biological or artificial, use hierarchical structures to predict the next word from context and thereby understand language.
As Goldstein’s study notes, the brain’s language processing is a continuous cycle of predicting and correcting errors based on context—essentially the same principle behind GPT’s “next‑word prediction.” When we listen to stories, the brain uses the past few hundred milliseconds to compute future probabilities, a computation that AI models simulate with mathematical precision.
Conclusion: the future of intelligence
Reexamining the 20‑watt brain and the elephant with 257 billion neurons shifts our focus from quantity to algorithm. Human intelligence rests on efficient compression of chaotic information and on outsourcing cognitive load to language and writing.
Today we build LLMs that consume vast power. Although current AI remains far less energy‑efficient than the human brain (which needs just 20 watts), these systems are emerging reasoning abilities through a strikingly similar hierarchical prediction mechanism. Understanding this connection helps us better grasp AI and, for the first time in silicon, clearly reveals how human intelligence itself operates—a grand epic of combating universal entropy at exceedingly low energy cost.
Related Articles:
Temporal structure of human language processing mirrors hierarchical layers of large language models
▶️ Enjoyed this article? Creating great content takes effort!
Subscribe to our YouTube channel:
AI Rosetta
Your likes, subscriptions and shares mean the world to us! 🙏
For more AI news, visit:
https://www.airosetta.com