Why your brain is 3 milion more times efficient than GPT-4
Claims that the human brain is “3 million times more efficient” than GPT‑4 prompt scrutiny of how energy use is being compared, with several commenters noting the article conflates training and inference costs and relies on rough, sometimes incorrect back-of-the-envelope math. Others argue that such comparisons are inherently muddled because human cognition is shaped by evolution, embodied learning, and hardwired structure, while LLMs are narrow text predictors running on very different hardware. The thread also branches into debate over whether current models genuinely “understand” language or are best seen as powerful but fallible conversational search tools, and whether it’s meaningful to equate machine and human intelligence at all.
Article quality and focus
- Many commenters found the title (“3 milion more times efficient…”) clickbaity and the spelling error distracting.
- Several felt the piece is a long, beginner-level “wall of text” heavy on basic computing (bits, ASCII) and light on the promised brain vs GPT-4 analysis.
- The vector database comparison is viewed by some as hand-wavy and even ad-like, lacking clear benchmarks, dataset descriptions, or rigorous methodology.
Energy efficiency: brain vs GPT-4
- Multiple comments argue the comparison mixes training energy for LLMs with inference energy for humans, which is not apples-to-apples.
- One rough recalculation (after correcting a kcal vs calorie error) suggests human brains may be only slightly more efficient than GPT-4 in “training” and less efficient during inference, under specific assumptions.
- Others note that if you include the energy cost of evolution, upbringing, education, or the infrastructure supporting humans, the accounting becomes extremely complex and somewhat arbitrary.
- There are nitpicks about misuse of units (e.g., “Watts per hour”) and simplistic analogies (one human vs entire GPT-4 data center).
Intelligence, understanding, and creativity
- Heated debate over whether LLMs “understand” language or merely do statistical prediction.
- Some insist only brains truly think, are original, and have qualia; LLMs are powerful “stochastic parrots.”
- Others argue humans are also pattern recognizers constrained by prior data, and that the distinction between memorization and understanding is blurry and methodologically unclear.
- Creativity is contested: one side claims humans can iteratively build genuinely novel concepts; the other says both humans and AIs just recombine existing patterns.
Practical value and limits of LLMs
- Several participants treat GPT-like systems as conversational search engines: good for summaries, code generation, and format transformations, but untrustworthy without verification.
- Some users report great success with programming help; others recount persistent hallucinations and factual errors (especially with certain models).
Brain vs computer models and hardware
- Commenters stress that brains and digital computers are fundamentally different physical systems; the “brain as computer” is a metaphor with limited reach.
- Points raised about the brain’s heavy “pretraining” via evolution and hardwired structure.
- Neuromorphic chips are mentioned as a promising direction for more brain-like, energy-efficient computation.