LLMs are as energy-efficient as brains

0 views
Skip to first unread message

Anand S

unread,
Apr 25, 2026, 11:03:48 AM (4 days ago) Apr 25
to s-a...@googlegroups.com

For a typical GDPVal style task, humans take ~7 hours and the brain consumes ~135 Wh.
Frontier LLM agents spend 50-500 Wh.
So, we may already be 3x more or less efficient than the brain.

Roughly in the same ballbark!

A colorful, hand-drawn sketchnote infographic comparing the energy efficiency of the human brain versus AI large language model (LLM) agents on complex knowledge tasks (GDPval-style). The vertical 9:16 layout uses bold comic-style lettering, doodles, and icons to explain that humans and AI are roughly comparable in energy use today. At the top, a smiling cartoon brain and a friendly robot stand beside a server rack under the title “Human Brain vs LLM Agents: Who uses less energy to solve hard tasks?” A highlighted “Big Takeaway” section states that AI agents are about 0.3× to 3× as energy-efficient as the human brain, and 2× to 10× more efficient if the whole human body is considered. A central panel explains the concept “Energy = Power × Time,” showing the brain as a low-power, long-duration “marathoner” and the AI as a higher-power, short-duration “sprinter.” The human section illustrates a person working for about 404 minutes (6.7 hours), with the brain consuming roughly 20 watts and about 135 watt-hours per task. The AI section shows a robot completing the same task in about 4.5 minutes, using an estimated 75,000 to 600,000 tokens and roughly 50 to 500 watt-hours of energy. A comparison panel uses a tug-of-war between the brain and robot to show that AI is within a single-digit factor of the brain in efficiency, not orders of magnitude better. Additional panels explain why estimates vary, including factors like whether only the brain or whole body is counted, token usage, and human review overhead. A final summary states that humans think slowly at low power, while AI thinks quickly at higher power, making them roughly comparable today. The infographic includes icons, speech bubbles, and labeled source references to OpenAI GDPval, GPT-5.4 reporting, brain energy use (~20 W), and public estimates of AI energy consumption.

Reply all
Reply to author
Forward
0 new messages