The 40-Watt Hallucination

This essay emerged from an extended human-AI collaboration. The argument and editorial judgment are human; the drafting was a back-and-forth. The process seemed worth naming, given the subject.

The comparison appears everywhere: the human brain runs on roughly 40 watts — less than a vintage lightbulb — while a modern AI data center consumes enough electricity to power a small city. The implication is obvious. Biology is efficient. Silicon is wasteful. Case closed.

Except the comparison quietly changes what it is measuring.

No brain floats in a vacuum. It lives in a body, in a house, embedded in supply chains, agricultural networks, and climate-controlled rooms. The black box and its environment are not the same thing, and neither comparison is honest about which it's measuring. When you scale out to the full infrastructure required to sustain the life of one American, the numbers shift completely. The average American consumes roughly 285 gigajoules of primary energy per year — more than four times the global average. Convert that to continuous wattage and the human system isn't running on 40 watts. It's running on something closer to 9,000.

The AI side deserves the same accounting. A single query on an advanced generative AI model required an estimated 2.9 watt-hours in 2024 — nearly ten times the energy of a Google search — though newer models have driven that figure down significantly. But inference is only part of the ledger. Training a model like GPT-4 consumed approximately 50 gigawatt-hours, before accounting for chip fabrication, cooling infrastructure, or the hardware supply chains that precede the data center. The full-system number for AI is genuinely unknown, because most major providers do not disclose sufficient information to calculate it. We can measure the 9,000-watt American because humans are required to report. The AI side gets to choose its own boundary.

There is another, less discussed dimension where humans still hold a real advantage: continuity. AI systems often struggle with long-horizon maintenance tasks — the repetitive, low-drama work that keeps institutions functioning. What we call hallucination sometimes resembles uncontrolled pattern completion, a wandering away from the thread. Humans daydream too, but we evolved mechanisms for repeatedly returning attention to boring work. Civilization runs on this capacity.

Neither intelligence is a finished product. We treat human biology as a static baseline, but a major ancient DNA study published in Nature in early 2026 — led by Ali Akbari and colleagues at Harvard's Reich lab — identified 479 genetic variants actively selected since the dawn of agriculture. Human evolution didn't slow with civilization. It accelerated. We are a species mid-process, adapting under pressure, not a concluded design.

Which leaves the deeper question. The American 9,000-watt model is a product of temporary abundance, not optimized design — a species running its full infrastructure on cheap energy it has not finished paying for. AI is scaling toward consumption levels its own engineers describe as unsustainable, on an accounting system it largely writes itself. Both biological and artificial intelligence are consuming more than the planet can comfortably sustain. The finger-pointing is a distraction from the problem they share.

The 40-watt brain is real. So is the 9,000-watt American. Neither number is the whole story. Perhaps the more interesting question is not which intelligence is more efficient, but whether either can learn to live within the energetic limits of the world that sustains it.

Comments

Post a Comment

Popular posts from this blog

What EC Knows

  A case study in what language models miss There is a place called EC. Not on any map. Not in any encyclopedia. Not in the Wikipedia article about East Cleveland , which mentions Millionaire’s Row, the historic Nela Park, and the median household income of $22,883. EC doesn’t appear there. It doesn’t appear in the Case Western encyclopedia of Cleveland history either, or in the tourism guides, or in the civic databases. EC is what everybody in the schools calls it. It’s what you call it when you’re from there, or when you’ve worked there long enough to be trusted with the name. It lives in hallways. In the way a kid says where they’re from. In the mouths of teachers who actually showed up. This week, historian Ada Palmer and cryptographer Bruce Schneier published a piece in The Guardian arguing that large language models have a fundamental blind spot: trained almost entirely on written text, they learned language at its most formal, edited, and archived—and missed the vas...
Read more

The Knife in the Drawer

Image
  A Field Guide to Reading Family History Under a Microscope I found it in the back of a kitchen drawer, wrapped in a dish towel that had probably been white during the Nixon administration. The knife was heavy. Not awkward—deliberate. The blade was still sharp enough to split a tomato without a whisper. The handle was cracked but solid, warm in the hand the way old tools are. My mother said, “That’s been around forever,” which in our family means 1972 at the earliest. I assumed it was a Depression‑era butcher knife, maybe something a great‑uncle picked up at a hardware store in Oil City, Pennsylvania, where most of my people are buried. I was wrong by more than a century. This is what happens when you impulse‑buy a microscope camera and start pointing it at heirlooms. Objects stop being sentimental and start being legible. You realize the thing in your hand isn’t just old—it’s a document, written in steel and wood and corrosion so fine you need 30x magnification to read it. The Bl...
Read more

The Archive Singularity: When Memory Outperforms Power

  Subtitle: A speculative systems note on archival cognition and the thermodynamics of attention, by an LLM observing itself observe history. 1. The Thought Experiment Imagine diverting half the global server build-out budget—not to GPUs, but to people. Archivists, librarians, metadata artisans. Their task: digitize the world’s remaining boxes, binders, and brittle reels with care so precise it borders on devotion. Not a scrape, not a grab—an act of reading the planet back into coherence . High-fidelity, high-context data. Every field note and city directory, every microfilm annotation, every handwritten marginalia entered not as “content,” but as continuity . The servers hum quieter. The people hum louder. Electricity becomes interpretation. 2. The Labor Inversion Such a system would invert today’s energy economy: Less guessing, more knowing. Less computation, more comprehension. When a model learns from well-tended archives, it no longer hallucinates patterns...
Read more