Can Intelligence Be Weighed? A Physicist Uses Thermodynamics to Put a Ruler on Intelligence
An Old Problem
In 1867, Maxwell imagined a little demon guarding a door between two gas chambers, letting fast molecules through while blocking slow ones. Soon one side got hot and the other cold, with no work done — the second law of thermodynamics seemed broken.
This "Maxwell's demon" troubled physicists for nearly a century, until Landauer and Bennett nailed it down: the demon must erase memories, and erasing memory itself dissipates heat. The demon is not free.
But in a June 2026 paper (arXiv:2606.20231), Ishanu Chattopadhyay recruits this old demon for something else — as a measuring stick.
His question is not "can the demon exist," but something harsher:
> If Maxwell's demon, large language models, human writers, thermostats, and a rock all count as "having some kind of intelligence," is there a unified, measurable physical quantity to compare them?
The answer is yes. The quantity is called rare-valid lift. It turns intelligence into a single thermodynamic number.
What Does Intelligence Actually "Do"
Forget brains and GPT for a moment. Ask a plainer question: what does a system — any system — actually do to the world?
It changes the probability distribution of the future.
A rock lying on the ground leaves the future distribution basically at "stays put." A thermostat senses temperature and controls a switch, twisting the future distribution — the event "room stays at 22 degrees" becomes more likely. A person writes a poem; the probability of "this string of text appearing on paper" is unimaginably higher than for a monkey on a keyboard.
But there's a trap. If I merely push probabilities toward futures that were already easy to happen, that isn't intelligence. Pushing a car that's already rolling downhill doesn't count — that's going with the flow.
Intelligence should mean: pushing probability toward futures that are unlikely but still valid.
- "Valid" is domain-dependent. In physical systems, valid means obeying the laws of physics. In biological systems, valid means surviving. In language systems, valid means grammatical, coherent, and factually correct.
- "Rare" is relative to a passive baseline — what the world would do if the system didn't intervene.
- Rock: P = P_0, lift is 0.
- Thermostat: pushes "room stays at temperature" from passive ~0.01 to ~0.9, a lift of about 89×.
- Maxwell's demon: pushes "spontaneous temperature difference" from exp(−ΔS/k_B) to nearly 1, a lift of roughly 10^3146.
- Turing test / Legg-Hutter: behavior-based, measuring performance across environments. "Task-facing."
- ARC / Chollet: based on skill-acquisition efficiency — how many samples you need to abstract a rule. Also task-facing.
- Rare-valid lift: path-facing. It doesn't ask "how well do you perform on a task," but "once the description level, baseline law, validity criterion, and observation resolution are fixed, what operation did you perform on the path probability distribution?"
- Free energy / active inference: free-energy minimization is a variational description of "how to perceive and act"; rare-valid lift is a measurable "how much were path probabilities reweighted." Complementary.
- Semantic information: "information is meaningful iff causally necessary for survival" is counterfactual; rare-valid lift is about path probabilities.
- Legg-Hutter: expected reward over an environment distribution vs. path-probability lift under a fixed baseline. "Average performance" vs "underlying operation."
The measure then emerges:
In plain words: over the set of rare-but-valid futures V_δ whose passive probability is only δ, how many times more likely does the system make them (P(V_δ)) compared to the passive probability P_0(V_δ).
Ten to the 3146th power. This number is too large for a log axis — the author had to use a double logarithm Λ = log₁₀(log₁₀(I+1)+1) to fit everything in one table.
Recursive Self-Simulation: The Necessary Architecture of Intelligence
A measure alone isn't enough. The key question: what kind of system can achieve this?
The author's answer: recursive self-simulation.
An intelligent system must carry an internal "world model," and that model must include itself. It must predict not only "if I take action a, how will the world change," but also "after I do a, how will my information state change, and what will I do then" — a self-referential recursion.
In Minsky's words: advanced problem-solving requires a system to represent its own goals, resources, and problem-solving activities, and self-understanding can involve "models of models of models."
Formally, system B maintains an internal representation hierarchy:
where U = B ∪ E is the totality of "system + environment." Since B ∈ U, a sufficiently general world model must also represent the system itself — its state, memory, uncertainty, actions, and future information updates. Recursion follows naturally.
This recursion isn't decoration; it's mathematically necessary. The paper's core theorem states:
> Theorem (necessity): Under finite amplification capacity, if the system's internal simulation cannot identify rare-valid futures with high fidelity, it cannot achieve high rare-valid lift.
> Theorem (near-sufficiency): When rare-valid fidelity is high enough and the simulation contains an effective policy, achievable lift approaches the optimum allowed by the actuation-capacity bound.
In plain language: you must first find the "rare but valid futures" in your head before you can push probability toward them in reality. Inaccurate identification means no push; accurate identification means you can push nearly to the limit.
This matches intuition. A chess player who cannot mentally simulate "if I move here, how will my opponent respond, and then what should I do" cannot play well. A scientist who cannot mentally run the loop "if hypothesis A holds, what will the experiment show, and how should I revise" cannot do good research.
Recursive self-simulation isn't a "feature" of intelligence — it's its engine.
A Ruler That Weighs Everything
The most exciting part is Table 2 — the intelligence scale of everything. Using Λ = log₁₀(log₁₀(I+1)+1), the author fits all systems from rocks to Maxwell's demon into one table:
| System | Λ value | Intuition | |------|------|------| | Passive matter (rock) | 0 | No intervention, P = P_0 | | Fixed feedback control (simple thermostat) | 0.114 – 0.477 | Amplification α = 2 to 100 | | Repetitive dynamic control (multi-level feedback) | 0.493 – 0.603 | 7–10 levels of binary improvement | | GPT-5 text generation | ≈ 1.170 | Sentence-level lift with entropy-rate correction | | Human text generation | ≈ 1.196 | Slightly above GPT-5 | | Maxwell's demon (single event) | 3.498 | ΔS = 10⁻¹⁹ J/K | | Maxwell's demon (velocity selection in 1 mm³ of air) | 14.653 – 15.360 | Aggregated selection over ~10¹⁶ particles |
Several details here are spine-chilling.
First, GPT-5 and human writers nearly tie on this scale, with humans slightly ahead (Λ 1.196 vs 1.170). The author uses Louwerse's composite estimate: there are about 5×10²¹ valid, interpretable sentence-level strings in English, while the "human-quality" target set is about 10⁷, so δ* ≈ 2×10⁻¹⁵ and human lift is about 5×10¹⁴. GPT-5 scores slightly lower because its entropy rate is a bit higher (0.74 vs 0.77 bits/char), giving it a larger typical set under the same quality constraint.
Second, the Maxwell demon's intelligence is 10^3146. This isn't a metaphor — it's a direct computation from the fluctuation theorem: the passive probability of a −ΔS entropy-decreasing event is exp(−ΔS/k_B); the demon pushes it to nearly 1, so the lift is exp(ΔS/k_B). With ΔS = 10⁻¹⁹ J/K and k_B = 1.38×10⁻²³ J/K, ΔS/k_B ≈ 7243, so I+1 ≈ e^7243 ≈ 10^3146.
Third, the velocity-selection demon in 1 mm³ of air reaches Λ ≈ 15, because it aggregates microscopic selections over ~2.44×10¹⁶ particles, each contributing a bit of rare-valid lift that multiplies up.
From a rock to 1 mm³ of demon-managed air, Λ spans 0 to 15 — fifteen orders of magnitude on a log-log scale. This ruler really can weigh everything.
Why This Ruler Is "Right"
You might ask: why is this definition better than the Turing test, Legg-Hutter intelligence, or the ARC benchmark?
The author is honest: the framework is not a replacement but a different perspective:
This distinction matters. Task-facing measures tell you the outcome; path-facing measures tell you the underlying operation. The latter enables cross-species, cross-substrate comparison — brains, LLMs, microbial communities, immune systems, controllers, and Maxwell's demon are incomparable on task-facing metrics (you can't make E. coli take ARC), but on path-facing metrics they all become comparable as "changes to future probability distributions."
The framework also connects with existing thermodynamic/information-theoretic concepts rather than conflicting:
Engineering Insights: What This Means for AI Practitioners
This paper isn't pure philosophy. It has concrete implications:
1. "Recursive self-simulation" is a testable architectural hypothesis. The paper proves high rare-valid lift requires high-fidelity rare-valid simulation. To assess an agent's "intelligence ceiling," don't just look at benchmark scores — look at how accurately its internal model identifies rare-valid futures. A new evaluation paradigm: measure "what it found in its head," not just "what it got right."
2. LLMs and humans nearly tie on this scale, but differently. GPT-5's Λ ≈ 1.170 vs human 1.196, but GPT-5's entropy rate is lower (0.74 vs 0.77) — a larger typical set, a broader range of valid strings. Humans are slightly higher because their target set under quality constraints is smaller and rarer. In short: LLMs are broader, humans are more precise.
3. Multi-level control multiplies. The repetitive dynamic control row (Λ = 0.493–0.603) shows 7–10 binary improvement levels of 2× each, totaling 2⁷ to 2¹⁰. Multi-level agent intelligence is the sum of per-level log-lifts — each level's ability to identify rare-valid futures contributes multiplicatively; one weak link collapses the chain.
4. Maxwell's demon is the "upper anchor" of intelligence. 10^3146 isn't fantasy; it follows directly from the fluctuation theorem. It's a physical ceiling: like the speed of light for velocity — unattainable, but telling you how far there is to go.
My Reflection: A "Mass-Like" Definition of Intelligence
Physics unified the concept of mass: Newtonian inertial mass, then Einstein's E=mc², then gravitational mass.
Intelligence measurement is undergoing something similar. The Turing test is "behavioral mass" — do you act intelligent? Legg-Hutter is "universal intelligence mass" — expected reward across computable environments. ARC is "learning-efficiency mass."
Chattopadhyay proposes "path-probability mass" — what operation you performed on the future distribution. Its ambition: not measuring on a task, environment, or benchmark, but at physics's lowest level — the path probability distribution. It pulls intelligence from "psychology" into "statistical mechanics."
There are limits, of course. Most obviously: "validity" is domain-dependent with no universal definition. Physical validity is the laws of physics; biological validity is survival; linguistic validity is grammar + semantics + facts. Your validity definition directly determines your Λ value. The author candidly frames this as "level-relativity" — intelligence is not an absolute value but relative to a chosen description level, baseline law, validity criterion, and observation resolution.
But that's precisely its strength. It doesn't pretend intelligence has an absolute value independent of context. It says: fix a context, and intelligence becomes a measurable physical quantity. Change the context, the number changes — but the measurement method doesn't.
That's more honest, and more physical, than any "universal intelligence test."
Code and Data
Numerical calibration code: https://github.com/zeroknowledgediscovery/tme
Data archived at Harvard Dataverse: https://doi.org/10.7910/DVN/F5TGT3
GPT-human entropy rate estimates: https://github.com/zeroknowledgediscovery/nero
Conclusion
Maxwell's demon was invented in 1867 to challenge the second law of thermodynamics. It failed — Landauer proved erasing memory dissipates heat; the demon is not free.
But 159 years later, it's been brought back as a ruler. This time it's not the villain — it's the ceiling of the scale. From rock to demon, Λ spans 0 to 15, with every intelligent system arranged along this axis.
And we, as human writers, sit at Λ ≈ 1.196. GPT-5 sits beside us at Λ ≈ 1.170 — a gap of 0.026 double-log units.
How big is that gap? Roughly an 8× difference in "rare-valid lift." Sounds small. But remember, this is under one specific context: sentence-level, English, human-quality threshold. Change the context — proving theorems, writing code, designing experiments — and the number changes.
Intelligence is not a number; it's a position on a ruler. And the ruler's markings are given by thermodynamics.
---
Paper: Chattopadhyay, I. *Thermodynamic Measure of Intelligence*. arXiv:2606.20231 (2026). Code: https://github.com/zeroknowledgediscovery/tme Categories: cs.AI, cond-mat.stat-mech, cs.IT, math-ph, nlin.AO