The floor under every thought
Erasing one bit costs at least kT·ln 2 of heat, about 2.9 × 10⁻²¹ joules at room temperature. Measured in 2012. Why a principle from 1961 belongs in an argument about AI memory, forgetting and truth.
1 Nov 2026 · 2 min readInformation is physical. That sentence sounds like philosophy. It is a measured result with a number attached.
The floor
Landauer’s principle: erasing one bit of information dissipates at least
E ≥ k_B · T · ln 2
At room temperature that is about 0.018 eV, or 2.9 × 10⁻²¹ joules 1. Tiny. Not zero.
In 2012 a team from Lyon, Augsburg and Kaiserslautern measured it: the heat released when a single bit, a colloidal particle in a double-well trap, was erased, approaching the Landauer bound 2. The floor is not a thought experiment any more.
How far above it we compute
As of 2012, ordinary computers used about a billion times the Landauer energy per operation 1. Almost all of that is not thinking. It is moving data and throwing information away carelessly.
| energy per bit operation | |
|---|---|
| Landauer floor, room temperature | 2.9 × 10⁻²¹ J |
| ordinary computers, 2012 | about 10⁹ × the floor |
That gap is the room architecture works in. It is why we keep saying the next large gains come from architecture before they come from bigger models.
Three things it changes in how we build
1. Forgetting is not free, and not optional. Every erasure costs heat, so a system that never forgets is not “lossless”: it has pushed the cost somewhere else, into a store that grows until nothing in it can be found. Our memory design treats forgetting as metabolism, the price of usability. The graph-memory literature reaches the same place from engineering: shard what you know, retrain only what a forgetting request touches 3.
2. Truth is a property of matter, not only of text. If information is physical, it depends on the substrate it runs on. A number on this site is a claim about a database, a log or a measurement on one machine. That is why every article names its record. Provenance is not decoration on a physical claim. It is the claim’s address.
3. Low entropy is the target, not maximum intelligence. Water reaching the valley is not clever. It is following the gradient to a lower energy state. The systems we want behave the same way: efficient, near the floor, and indifferent to looking smart.
Where this stops
Landauer bounds erasure, not computation as such. Reversible computing can in principle compute without erasing and so below this floor. And nothing here says how much of an AI system’s energy is avoidable in practice; the billion-fold figure is a 2012 statement about general-purpose computers, not a measurement of any model.
What it does give us is a direction. There is a floor, it has been measured, and everything above it is design.
- Landauer's principle. Wikipediasaved pagecorpus: ~/arch-of-truth
- A. Bérut, A. Arakelyan, A. Petrosyan, S. Ciliberto, R. Dillenschneider, E. Lutz. Experimental verification of Landauer's principle linking information and thermodynamics. Nature 483, 187–190 (2012)not in corpus
- J. Miao, L. Hu, Q. Zhang, L. Cao. Graph Memory Learning: Imitating Lifelong Remembering and Forgetting of Brain Networks. arXiv:2407.19183 (2024)originalcorpus: ~/phyto
We clustered 1,046 knowledge-graph entities and found nothing, at every k, with two methods. Two clean hypotheses for why were falsified. The structure was there all along, in a different representation.
A spruce that grew from a cold seed buds earlier, decades later. The DNA is the same; how it is read is not. What that taught us about our own memory, including the bug where we got it backwards.
One day of mistakes, read back. Every one came from reasoning like an intelligence. Every correct step came from reading what was there and following the gradient.