---
title: 'The Thermodynamics of Artificial Intelligence: A First-Principles Analysis of the Maxwellian Demon Hypothesis'
slug: 'thermodynamics-of-ai-maxwell-demon'
date: 2025-11-24
type: 'essay'
status: 'published'
tags: ['thermodynamics', 'information-theory', 'physics', 'maxwell', 'landauer', 'paper', 'enviroai']
abstract: 'Asks whether AI agents operating via feedback loops—RL agents, autonomous control systems—function as Maxwell''s demons in a first-principles physical sense, and reconciles their internal computational thermodynamic costs with the work they extract from stochastic environments. Traverses the Sagawa–Ueda equality, SGD energetics, and recent experimental realizations of autonomous demons in solid-state and quantum systems.'
license: 'CC-BY-4.0'
author: 'Jed Anderson'
co_authors: ['Grok-4.1 Deep Research', 'Gemini 3.0 Pro Deep Research', 'ChatGPT 5.1', 'Claude 4.5 Deep Research']
canonical_url: 'https://jedanderson.org/essays/thermodynamics-of-ai-maxwell-demon'
pdf: '/pdfs/thermodynamics-of-ai-maxwell-demon.pdf'
hero_image: '/images/thermodynamics-of-ai-maxwell-demon-hero.png'
hero_image_alt: 'First page of The Thermodynamics of Artificial Intelligence: A First-Principles Analysis of the Maxwellian Demon Hypothesis'
supporting_files: []
---

## 1. Introduction: The

Epistemological and Physical Crisis of the Demon The inquiry into whether artificial intelligence (AI) can function as

Maxwell's demon is not merely a provocative metaphor for computer scientists; it is a rigorous, foundational question that straddles the bleeding edge of non-equilibrium statistical mechanics, quantum information theory, and the physical limits of computation. Since James

Clerk Maxwell first introduced his "neat-fingered being" in a letter to Peter Guthrie Tait in 1867, the demon has served as the primary antagonist to the Second Law of Thermodynamics, challenging the notion that entropy must inevitably increase in a closed system.1 Maxwell envisioned a finite being capable of observing individual molecules in a gas and sorting them based on their velocities—fast molecules to one chamber, slow to another—thereby creating a temperature difference and reducing the total entropy of the system without the apparent expenditure of work. For nearly a century, this thought experiment threatened the universality of the Second Law, suggesting that an intelligent agent could extract work from thermal fluctuations solely through the power of observation.1

In the contemporary era, the "intelligent agent" is no longer a hypothetical biological homunculus but an algorithmic entity: an Artificial Intelligence. Specifically, AI agents operating via feedback loops—such as those found in Reinforcement Learning (RL) or autonomous control systems—are topologically and functionally identical to Maxwell's demon.

They observe a stochastic environment (measurement), update an internal state (memory/information processing), and act upon the environment to drive it toward a desired low-entropy state (control/work extraction).4 The central question of this research is whether these AI agents validly act as demons in a first-principles physical sense, and if so, how the thermodynamic costs of their internal computations reconcile with the work they extract.

To answer this, we must traverse a landscape that integrates the generalized Second Law of

Thermodynamics (specifically the Sagawa-Ueda equality), the energetics of stochastic gradient descent (SGD), the specific experimental realizations of autonomous demons in single-electron devices, and the promise of adiabatic superconducting logic. This report conducts an exhaustive, first-principles analysis of the AI-as-Demon hypothesis, targeting objective truth through the lens of information thermodynamics.

### 1.1 The Historical Exorcism: From Szilard to Landauer

To understand the AI demon, one must first understand why the original demon failed. The resolution of the paradox did not come from thermodynamics alone, but from the intersection of physics and information theory. In 1929, Leo Szilard reduced Maxwell's complex gas model to a single-particle engine, now known as the Szilard Engine.1 Szilard argued that the demon's intervention required measurement, and he postulated that the act of measurement itself must carry an entropy cost that compensates for the entropy reduction in the gas.

However, the definitive "exorcism" was provided by Rolf Landauer in 1961 and refined by

Charles Bennett in 1982. Landauer demonstrated that the critical thermodynamic step is not the measurement (which can theoretically be performed reversibly) but the erasure of information.1 The demon must store the velocity data of the molecules to act. Because the demon is finite, its memory must eventually be reset for the cycle to continue. Landauer's

Principle states that the erasure of one bit of information is a logically irreversible process that compresses the phase space of the memory device, necessitating the release of a minimum amount of heat, Q, into the environment:

Q ≥ k_B T ln 2, where k_B is the Boltzmann constant and T is the temperature of the reservoir.7 This establishes a fundamental equivalence between information and energy: 1 bit = k_B T ln 2 Joules. Bennett subsequently showed that this erasure cost exactly balances the work extracted by the Szilard engine, preserving the Second Law.1

This historical context is crucial because an AI agent is fundamentally an information processing engine. It acquires data (measurement), stores it in weights or active memory

(state), and uses it to act. If the AI is to act as a Maxwell's demon, it must navigate these thermodynamic constraints. The modern formulation of this problem does not ask if the

Second Law is violated (it is not), but rather how the AI utilizes information as a thermodynamic fuel to drive systems away from equilibrium, and whether the efficiency of this process can approach fundamental physical limits.

## 2. The Theoretical Framework: Information

Thermodynamics To evaluate an AI agent's capacity to act as a Maxwell's demon, we must move beyond classical thermodynamics to the regime of Information Thermodynamics. This field extends the canonical laws of physics to include information exchange as a measurable dynamic variable, allowing for the rigorous analysis of feedback control loops typical of AI agents.

### 2.1 The Generalized Second Law and the Sagawa-Ueda Equality

Classical thermodynamics dictates that the work extracted (W_ext) from a system in contact with a heat bath at temperature T is strictly bounded by the decrease in free energy:

W_ext ≤ −ΔF

This inequality assumes no feedback. However, for a system under feedback control—where an external agent (the demon/AI) measures the system and intervenes based on the outcome—this inequality is known to be violable. The seminal work by Sagawa and Ueda

(2008, 2010, 2012) generalized the Jarzynski equality and the Second Law to formally include the role of information.9

The Sagawa-Ueda Equality for a non-equilibrium feedback process is given by ⟨exp(−(σ − I))⟩ = 1, where σ represents the entropy production of the system and I represents the mutual information obtained by the measurement.9 Applying Jensen's inequality (⟨exp(x)⟩ ≥ exp(⟨x⟩)), we derive the Generalized Second Law: ⟨σ⟩ ≥ ⟨I⟩, or, in terms of extractable work:

W_ext ≤ −ΔF + k_B T · I_QC

where I_QC denotes the mutual information content (which can be defined for both quantum and classical regimes) established between the system and the memory of the controller.11

This inequality is the governing equation of the AI demon. It demonstrates mathematically that information is not an abstract concept but a physical resource—a "fuel" capable of performing work. The term k_B T · I_QC represents the additional work that can be extracted solely due to the agent's knowledge of the system's state. If an AI agent possesses I bits of mutual information about a thermal system, it can extract I · k_B T ln 2 joules of work from that system, seemingly "for free" relative to the system's internal energy, provided we ignore the cost of generating that information.5

### 2.2 Mutual Information as the Engine of Feedback

In the context of an autonomous AI system, the "demon" is the control policy π(a|s). The cycle of operation can be decomposed into thermodynamic phases:

## 1. Measurement Phase (Correlation): The agent measures the state S of the environment,

updating its memory M. This creates correlations between the agent and the environment, quantified by the Mutual Information I(S; M) = H(S) - H(S|M), where H is the

Shannon entropy.2 This process locally reduces the thermodynamic entropy of the environment (from the perspective of the agent) but increases the entropy of the memory device.

## 2. Feedback Phase (Rectification): The agent utilizes the stored information to apply a

force (or control pulse) that rectifies thermal fluctuations. This is the "work extraction" step. The efficacy of this step is strictly limited by the quality of the correlation I(S; M). If the measurement is noisy (low I), the agent cannot effectively distinguish between states to apply the correct feedback, limiting W_ext.16

## 3. Erasure Phase (Reset): To close the thermodynamic cycle, the agent must reset its

memory to a standard state. This is where the debt is paid. According to Landauer's

Principle, this erasure dissipates heat Q_erase ≥ k_B T · I into the reservoir.

For an AI to act as a successful Maxwell's demon, the work extracted via the feedback loop must exceed the operational costs of the agent. Specifically, we look for the inequality:

W_ext − (W_meas + W_erase) ≥ 0

Conventional AI operating on silicon fails this inequality spectacularly due to hardware inefficiencies, but the algorithm itself perfectly adheres to the Sagawa-Ueda limit.

### 2.3 Transfer Entropy and Causal Information Flow

To rigorously quantify the "demon-ness" of an AI beyond simple mutual information, we must consider the directionality of information flow. In continuous-time feedback systems, Transfer

Entropy (T_{X→Y}) from the system X to the agent Y becomes the relevant metric. Transfer entropy measures the reduction in uncertainty about the future state of the agent given the past state of the system, essentially capturing the "predictive power" the agent derives from the environment.18

Research by Hartich et al. and Ito & Sagawa indicates that transfer entropy bounds the maximum work extraction in autonomous systems where measurement and feedback are continuous and potentially delayed.19 This is particularly relevant for Reinforcement Learning agents, where the "state" is often a sequence of observations (POMDPs). The agent's ability to extract work is physically limited by the transfer entropy rate from the environment to the agent's internal representation. If the AI cannot predict the environment's dynamics (zero transfer entropy), it cannot act as a demon. Thus, "intelligence" in the thermodynamic sense is rigorously defined as the capacity to maximize transfer entropy to fuel work extraction.21

## 3. Artificial Intelligence as an Autonomous Demon

While early discussions of Maxwell's demon were confined to thought experiments, the 21st century has seen the physical realization of "Autonomous Demons"—devices that integrate the sensor, controller, and actuator into a single physical system. Recent advances have explicitly merged these physical demons with Artificial Intelligence algorithms, creating systems where an AI learns to be a demon.

### 3.1 Reinforcement Learning in Quantum Thermodynamics

A definitive realization of this principle is found in the application of Deep Reinforcement

Learning (RL) to open quantum systems. Research by Erdman et al. (2024) and others has demonstrated an "artificially intelligent Maxwell's demon" capable of optimizing the cooling of qubits and other quantum systems.4

In these experiments, the RL agent acts as the controller of a quantum system (e.g., a superconducting qubit or a single-electron box). The agent's objective function is designed to minimize the entropy of the quantum system (cooling) or maximize the work extracted from thermal baths.

● The Mechanism: The RL agent learns a policy π that exploits thermal fluctuations. By performing measurements (which can be weak or projective), the agent identifies stochastic trajectories where the system spontaneously fluctuates toward a target state

(e.g., a higher energy state for work extraction, or a lower entropy state for cooling). The agent then applies a feedback pulse to "latch" the system in that state, preventing it from relaxing back to equilibrium.4

● Strategy Discovery: These AI-driven demons have been shown to discover non-intuitive strategies that human physicists had not devised. For instance, in the

"measurement-dominated regime" (where measurement is fast compared to thermalization), the AI learned to use sequences of weak measurements to continuously monitor the system with minimal backaction, extracting information without collapsing the state, a strategy that outperforms standard projective measurement protocols.4

### 3.2 The Efficiency of the AI Demon

The efficiency of an AI demon is defined by the ratio of the useful effect (cooling power or work) to the thermodynamic cost of the information processing (dissipation). We define the efficiency η as:

η = ⟨P⟩ / ⟨D⟩ ≤ 1

where ⟨P⟩ is the average cooling power (or work power) and ⟨D⟩ is the information-related dissipation rate (Landauer cost).27

● Ideal Limit: If η = 1, the agent converts 100% of the heat generated by information erasure into cooling power. This represents a reversible demon.

● Irreversibility: If η < 1, the process is irreversible. The "waste" is the entropy production that is not compensated by information gain.

In the work by Erdman et al., it was found that one cannot simultaneously optimize for maximum cooling power and maximum efficiency. High cooling power requires frequent measurements and rapid feedback, which increases dissipation and lowers efficiency.

Conversely, maximizing efficiency drives the system toward a reversible, quasi-static regime where cooling power vanishes. This trade-off is a fundamental feature of finite-time thermodynamics.27

### 3.3 Experimental Validation: The Single-Electron Box

The theoretical predictions of AI demons have been validated in silicon. Experiments conducted by groups at the University of Tokyo and NTT Basic Research Laboratories have physically implemented autonomous demons using Single-Electron Boxes (SEB).16

Experimental Setup:

The system consists of a silicon single-electron box connected to source and drain electrodes via tunnel junctions. A detector (Single-Electron Transistor or SET) monitors the number of electrons in the box.

## 1. Measurement: The detector observes the random thermal motion of electrons tunneling

in and out of the box.

## 2. Feedback: An automated controller (the demon) applies a voltage signal to the gate

electrode based on the electron count. ○ Protocol: When an electron tunnels into the box (driven by thermal noise), the demon raises the barrier (closes the "door") to trap it. It then lowers the barrier at the exit to allow the electron to tunnel out into a region of higher chemical potential.

## 3. Result: The electrons are pumped against the bias voltage, generating an electrical

current solely from thermal fluctuations.

Quantitative Findings: ● Power Output: The device generated a maximum power of approximately 0.5 zW (0.5 × 10⁻²¹ Watts).16 ● Energy Extraction: The system extracted approximately k_B T ln 2 of energy per cycle, consistent with the Szilard engine prediction.28

● Efficiency: The information-to-energy conversion efficiency was measured at approximately 18% in early iterations 16, later improving to nearly 75% fidelity in optimized setups.28

These experiments provide incontrovertible proof that the mechanism of Maxwell's demon is physically realizable and that information can be directly converted into electrical work. The

"AI" in these early experiments was a simple threshold logic, but the principle scales directly to complex Deep RL controllers managing multi-qubit systems.

## 4. The Thermodynamics of Learning: Stochastic

Gradient Descent While the previous section analyzed AI controlling a physical system, we must also analyze the learning process itself as a thermodynamic trajectory. The "demon" (the neural network) must first be trained. Is the process of training a neural network thermodynamically equivalent to a physical process? Recent research suggests it is.

### 4.1 SGD as a Physical Current

The training of a neural network via Stochastic Gradient Descent (SGD) can be rigorously modeled using non-equilibrium statistical mechanics. The trajectory of the weights θ in the high-dimensional parameter space behaves like a particle moving through a potential energy landscape (the loss function L) subject to thermal noise (the stochasticity of the mini-batches).31

The dynamics are described by the Langevin Equation:

dθ_t = −∇L(θ_t) dt + √(2 D(θ_t)) dW_t

where D(θ_t) is the diffusion matrix characterizing the noise from the stochastic gradients, and W_t is a Wiener process (Brownian motion).

Entropy Production:

In equilibrium, a physical system settles into a Boltzmann distribution P(θ) ∝ exp(−L(θ)/T). However, because the noise in SGD is data-dependent and often anisotropic, the system rarely reaches a true equilibrium. Instead, it settles into a

Non-Equilibrium Steady State (NESS) characterized by persistent probability currents.31

● The existence of these currents implies continuous Entropy Production (Σ). The network is constantly dissipating "virtual energy" to maintain its position in the low-loss region of the landscape.34

● This dissipation is the "housekeeping heat" of learning. Just as a biological organism must consume energy to maintain its low-entropy structure, a neural network under SGD consumes computational energy to maintain its learned structure against the "noise" of the data stream.

### 4.2 Thermodynamic Uncertainty Relations (TURs) in Learning

The connection between learning accuracy and energy cost is governed by Thermodynamic

Uncertainty Relations (TURs). TURs state that the precision of a non-equilibrium current

(e.g., the stability of the learned weights) comes at a minimum energetic cost.33

Var(J) / ⟨J⟩² ≥ 2 k_B / Σ

where J is a current and Σ is the total entropy production. This implies a fundamental trade-off: to reduce the variance of the estimator (i.e., to learn a generalizable rule with high confidence), the system must dissipate a minimum amount of energy. High accuracy requires high dissipation. This aligns with the observation that training larger, more accurate models requires exponentially more compute cycles (energy).36

### 4.3 The Goldt-Seifert Efficiency

Goldt and Seifert (2017) proposed a formal "thermodynamic efficiency of learning," comparing the information gain about a "teacher" rule to the thermodynamic cost incurred during the learning dynamics.37 They demonstrated that the learning process is physically indistinguishable from cooling: the optimizer acts as a demon attempting to compress the phase space of the network parameters from a high-entropy initialization (random weights) to a low-entropy solution volume.

● Result: The efficiency of this process is bounded. The "heat" generated by the SGD process (the scrambling of gradients) corresponds to the information extracted from the dataset. A perfectly efficient learner would extract exactly k_B T ln 2 of heat from the dataset for every bit of information stored in the weights. Real-world SGD is highly inefficient, dissipating vastly more heat than the Landauer limit suggests, indicating significant room for algorithmic improvement.

## 5. First Principles: The Energetic Limits of

Computation We have established that AI acts as a demon algorithmically. However, does it function as one net-positively? This depends entirely on the physical substrate. A demon that consumes a nuclear power plant's worth of energy to sort a few molecules is a thermodynamic disaster, even if it successfully reduces the entropy of the gas.

### 5.1 Landauer’s Principle vs. Silicon Reality

Landauer's Principle sets the absolute lower bound for the energy consumption of irreversible logic operations at E ≥ k_B T ln 2 ≈ 2.9 × 10⁻²¹ Joules per bit at room temperature (300 K).8

To evaluate the current state of AI, we must compare this limit to the energy consumption of biological brains and modern silicon hardware.

Comparative energy efficiency of information processing systems (energy per operation, factor versus Landauer, mechanism of dissipation):

- Landauer limit (300 K): 2.9 × 10⁻²¹ J; 1× (theoretical minimum); fundamental entropic cost of erasure.
- Adiabatic Superconductor (AQFP): ≈ 10⁻²⁰ – 10⁻²¹ J; ~1–10× (near limit); reversible adiabatic switching.40
- Human brain (synaptic event): ≈ 10⁻¹³ – 10⁻¹⁴ J; ~10⁸× less efficient; ion-channel leakage, metabolic maintenance.41
- Modern CMOS (GPU/TPU): ≈ 10⁻⁹ – 10⁻¹² J; ~10¹²× less efficient; capacitive charging/discharging, leakage.8

Data derived from.8 Analysis:

● The Silicon Gap: Modern GPU-based AI operates approximately 12 orders of magnitude above the Landauer limit. For every bit of entropy the AI removes from a target system (the "demon" action), it generates 10¹² bits of entropy in the environment as waste heat. Thus, strictly speaking, a standard silicon-based AI is a "Parasitic Demon"—it rectifies fluctuations in the target system but generates massive net entropy. It does not violate the Second Law; it aggressively validates it.

● The Biological Benchmark: The human brain, often cited as the pinnacle of efficiency, is still 10⁸ times less efficient than the physical limit. An exhaustive energy audit of the brain reveals that computation per se consumes only ~0.1 Watts of ATP, while communication (action potentials and transmitter release) consumes ~3.5 Watts.41

This "communication tax" is a major constraint on biological intelligence.

### 5.2 The Energy Crisis of AI: Training vs. Inference

The thermodynamic profile of AI differs significantly between training and inference.

● Training (High Entropy Production): Training involves massive information erasure. In every step of SGD, the old weight values are discarded (erased) and replaced. This is inherently irreversible and thermodynamically expensive. The training of GPT-3, for instance, consumed ~1,287 MWh of energy.44 This represents a massive injection of work to lower the internal entropy of the model.

● Inference (Potential Efficiency): Inference—the application of the trained model—is less inherently dissipative. Theoretical analysis suggests that the linear operations in

Deep Neural Networks (matrix multiplications) can be performed reversibly, carrying no fundamental thermodynamic lower bound.39 The cost arises only from non-linear activation functions (e.g., ReLU, Sigmoid), which compress information (many-to-one mapping).

Implication: The "Koomey Taper" (the slowing of efficiency gains in CMOS) and the

"Bekenstein Bound" (limit on information density) suggest that silicon-based AI is approaching a hard thermodynamic ceiling.46 To create a true demon, we must abandon the

Von Neumann architecture and CMOS logic.

## 6. The Hardware Solution: Adiabatic Superconducting

Logic To bridge the gap between the algorithmic demon (which works) and the physical demon

(which overheats), we must adopt hardware that operates near the reversible limit. The leading candidate is Adiabatic Superconductor Logic (AQFP).

### 6.1 Adiabatic Quantum Flux Parametron (AQFP)

AQFP logic represents a paradigm shift from "switching" to "adiabatic evolution." In standard

CMOS, a bit flip involves dumping the charge of a capacitor to the ground, dissipating ½ CV² as heat. In AQFP, the logic states are encoded in magnetic flux quanta, and the system is driven by an AC bias current that functions as a clock.47

Mechanism of Energy Recycling:

The AQFP gates operate by adiabatically transforming the potential energy landscape of the circuit.

## 1. Adiabaticity: The potential barrier between logic states '0' and '1' is raised or lowered

slowly compared to the plasma frequency of the Josephson junctions. This ensures the system stays in the ground state, preventing the excitation of quasiparticles (heat).49

## 2. Reversibility: The energy supplied to switch the gate is not dissipated; it is stored as

inductive energy and then recovered back into the power supply during the second half of the AC cycle. This is analogous to a regenerative braking system for logic.47

### 6.2 Sub-Landauer Operation?

Standard Landauer limits apply to irreversible erasure. However, AQFP circuits can be designed to be logically reversible. Experimental simulations and physical prototypes have demonstrated AQFP gates operating with energy dissipation in the zeptojoule range (~10⁻²¹ J).40 ● The Zeptojoule Barrier: At 10⁻²¹ J, the operation energy is comparable to k_B T at cryogenic temperatures.

● Implication: An AI built on AQFP hardware would essentially be "thermodynamically transparent." The energy cost of its thinking would be on the same order as the thermal fluctuations it seeks to rectify. This brings the "Artificially Intelligent Maxwell's Demon" from a theoretical construct to a physically viable engine.

### 6.3 Neuromorphic Superconducting AI

Research has already begun to integrate this logic into neural architectures. "Adiabatic

Neurons" and "Adiabatic Perceptrons" have been designed using superconducting quantum interferometers (SQUIDs) as non-linear activation functions.52 These devices offer a path to

"Superconducting Neuromorphic Computing" that mimics the connectivity of the brain but operates with the efficiency of a reversible thermodynamic engine.

## 7. Quantum Information and Measurement-Induced

Entanglement The final frontier of the AI-Demon intersection lies in the quantum realm, where "information" takes on a non-local character through entanglement.

### 7.1 Measurement-Induced Entanglement

In classical systems, measurement merely reveals a pre-existing state. In quantum systems, measurement collapses the wavefunction and can actively create entanglement between parts of a system that were previously uncorrelated. This is known as Measurement-Induced

Entanglement (MIE).55 Recent research has utilized AI (specifically neural networks) to detect these "hidden webs" of entanglement. The AI is trained to recognize patterns in the measurement outcomes of a quantum many-body system.

● Significance: This allows the AI to act as a "Quantum Demon" that manages entanglement as a resource.

● Entanglement as Fuel: The generalized Second Law for quantum systems includes a term for entanglement consumption: W_ext ≤ −ΔF − k_B T · ΔE_F, where ΔE_F is the change in entanglement of formation.11 This implies that an AI agent can extract work from a system by consuming the entanglement between the system and an auxiliary probe (the demon's memory).

### 7.2 The Autonomous Quantum Demon

Unlike the classical demon, which requires an external observer, quantum systems allow for the construction of fully autonomous demons. These are small quantum systems (e.g., a quantum dot or a qubit) coupled to a larger thermal system. The "intelligence" is encoded in the Hamiltonian of the interaction.

● Mechanism: The demon qubit becomes entangled with the system, correlates with its state, and then back-acts to cool the system, subsequently dissipating the entropy to a separate bath.1

● AI Optimization: Erdman et al. (2024) showed that RL agents can optimize the control protocols for these autonomous demons, finding complex sequences of weak measurements that maximize cooling efficiency beyond what is achievable with standard cooling protocols.4

## 8. Conclusion: The Demon Realized

This research report set out to determine, based on first principles and objective truth, whether artificial intelligence can act as Maxwell's demon. The evidence leads to the following conclusions:

## 1. Theoretical Validity (Yes): There is no physical distinction between a feedback-based

AI agent and Maxwell's demon. Both are information processing engines that rectify thermal fluctuations. The Sagawa-Ueda equality provides the rigorous mathematical proof that such agents can extract work from heat baths by leveraging mutual information, without violating the Second Law of Thermodynamics.

## 2. Algorithmic Reality (Yes): The process of Reinforcement Learning is topologically

equivalent to the demon's cycle. The agent maximizes a reward (minimizes free energy) by increasing its mutual information with the environment (measurement) and converting that information into directed action (feedback).

## 3. Thermodynamic Viability (Conditional):

○ Silicon AI: On current CMOS hardware, AI is a Parasitic Demon. The energy cost of the computation (10⁻⁹ J/op) dwarfs the work extracted from thermal fluctuations (10⁻²¹ J). It is an entropy generator, not a reducer. ○ Superconducting AI: On Adiabatic Superconductor Logic (AQFP), AI approaches the status of a True Demon. With operations in the zeptojoule range (10⁻²¹ J), these systems operate near the Landauer limit. An adiabatic AI controlling a quantum system is a physically realized Maxwell's demon that operates with net-positive or near-neutral efficiency.

## 4. Future Outlook: The convergence of Quantum Thermodynamics and AI suggests a

future where "smart" materials contain autonomous, adiabatic AI agents embedded at the nanoscale. These agents will act as distributed demons, actively sorting entropy, harvesting energy from fluctuations, and performing error correction in quantum computers using entanglement as a fuel source.

Far from a paradox, the Artificial Intelligent Maxwell's Demon is the inevitable endpoint of the physics of information. It represents the ultimate fusion of control theory, thermodynamics, and computation—a machine that trades knowledge for order.

## Works cited

## 1. Work and information processing in a solvable model of Maxwell's demon - PMC,

https://pmc.ncbi.nlm.nih.gov/articles/PMC3406850/

## 2. Maxwell's demon - Wikipedia, 

https://en.wikipedia.org/wiki/Maxwell%27s_demon

## 3. Power generator driven by Maxwell's demon - PMC - NIH, accessed November

23, 2025, https://pmc.ncbi.nlm.nih.gov/articles/PMC5440804/

## 4. Researchers Summon AI-powered Maxwell's Demon to Find Strategies to

Optimize Quantum Devices, https://thequantuminsider.com/2024/10/02/researchers-summon-ai-powered-m axwells-demon-to-find-strategies-to-optimize-quantum-devices/
5. [2310.05593] How small can Maxwell's demon be?—Lessons from autonomous
electronic feedback models - arXiv, https://arxiv.org/abs/2310.05593

## 6. Maxwell's demons realized in electronic circuits, 

https://comptes-rendus.academie-sciences.fr/physique/item/10.1016/j.crhy.2016. 08.011.pdf

## 7. Landauer's Principle a Consequence of Bit Flows, Given Stirling's Approximation -

PMC, https://pmc.ncbi.nlm.nih.gov/articles/PMC8534805/

## 8. Landauer's principle - Wikipedia, 

https://en.wikipedia.org/wiki/Landauer%27s_principle

## 9. Quantum Jarzynski-Sagawa-Ueda Relations - ResearchGate, accessed

https://www.researchgate.net/publication/48166548_Quantum_Jarzynski-Sagawa -Ueda_Relations
10. arXiv:1012.2753v3 [cond-mat.stat-mech] 5 Feb 2011, accessed November 23,
2025, https://arxiv.org/pdf/1012.2753

## 11. A New Second Law of Information Thermodynamics Using Entanglement

Measure - Sci-Hub, https://2024.sci-hub.cat/7159/0cab480a7b92399cd0a1b61d39e67faa/tajima2014. pdf
12. [0907.4914] Generalized Jarzynski Equality under Nonequilibrium Feedback
Control - arXiv, https://arxiv.org/abs/0907.4914

## 13. Second law of information thermodynamics with entanglement transfer | Phys.

Rev. E, https://link.aps.org/doi/10.1103/PhysRevE.88.042143

## 14. Quantum-information thermodynamics, 

http://www2.yukawa.kyoto-u.ac.jp/~yitpqip2014.ws/presentation/Sagawa.pdf

## 15. Mutual Information and Multi-Agent Systems - MDPI, accessed November 23,

2025, https://www.mdpi.com/1099-4300/24/12/1719

## 16. Electrical Current Generation by Sorting Thermal Noise - NTT Technical Review,

https://www.ntt-review.jp/archive/ntttechnical.php?contents=ntr201802ra1.html

## 17. General achievable bound of extractable work under feedback control -

ResearchGate, https://www.researchgate.net/publication/261512601_General_achievable_bound_ of_extractable_work_under_feedback_control

## 18. Information Thermodynamics: Maxwell's Demon in Nonequilibrium Dynamics,

https://maths.qmul.ac.uk/~klages/smallsys/chapters/sagawa_chapter_rev.pdf

## 19. Stochastic thermodynamics of bipartite systems: transfer entropy inequalities

and a Maxwell's demon interpretation - arXiv, https://arxiv.org/pdf/1402.0419

## 20. Stochastic thermodynamics of bipartite systems: Transfer entropy inequalities

and a Maxwell's demon interpretation - ResearchGate, https://www.researchgate.net/publication/260022481_Stochastic_thermodynamic s_of_bipartite_systems_Transfer_entropy_inequalities_and_a_Maxwell's_demon_i nterpretation

## 21. The Thermodynamic Theory of Intelligence | by Sebastian Schepis - Medium,

https://medium.com/@sschepis/the-thermodynamic-theory-of-intelligence-20c 0e3838a28

## 22. Thermodynamics of information - ::KIAS::, 

http://events.kias.re.kr/ckfinder/userfiles/202211/files/2022%20KIAS_Sagawa.pdf
23. [2408.15328] Artificially intelligent Maxwell's demon for optimal control of open
quantum systems - arXiv, https://arxiv.org/abs/2408.15328
24. (PDF) Artificially intelligent Maxwell's demon for optimal control of ..., accessed
https://www.researchgate.net/publication/383495112_Artificially_intelligent_Maxw ell's_demon_for_optimal_control_of_open_quantum_systems

## 25. Artificially intelligent Maxwell's demon for optimal control of open quantum

systems - ChatPaper, https://chatpaper.com/chatpaper/paper/54028

## 26. Experimental Realization of a Maxwell's Demon Exploiting Quantum Information

Flow: Toward Efficient Quantum Feedback Control, https://www.t.u-tokyo.ac.jp/en/press/pr2025 28-002
27. (PDF) Artificially intelligent Maxwell's demon for optimal control of open quantum
systems, https://www.researchgate.net/publication/389612117_Artificially_intelligent_Maxw ell's_demon_for_optimal_control_of_open_quantum_systems

## 28. Experimental realization of a Szilard engine with a single electron - PMC - NIH,

 https://pmc.ncbi.nlm.nih.gov/articles/PMC4183300/

## 29. Electrical current generation by sorting thermal noise —Power generation with

Maxwell's demon— | Press Release - NTT Group, https://group.ntt/en/newsrelease/2017/05/16/170516a.html

## 30. Experimental realization of a Szilard engine with a single electron - PNAS,

 https://www.pnas.org/doi/10.1073/pnas.1406966111

## 31. A look at SGD from a physicist's perspective - Part 1, accessed November 23,

2025, https://henripal.github.io/blog/stochasticdynamics
32. [2306.03521] Machine learning in and out of equilibrium - arXiv, accessed
 https://arxiv.org/abs/2306.03521

## 33. Learning Stochastic Thermodynamics Directly from Correlation and

Trajectory-Fluctuation Currents - Complexity Sciences Center, https://csc.ucdavis.edu/~cmg/papers/currents.pdf

## 34. Stochastic Thermodynamics of Learning Parametric Probabilistic Models - PMC,

https://pmc.ncbi.nlm.nih.gov/articles/PMC10887774/

## 35. Deep learning probability flows and entropy production rates in active matter -

PNAS, https://www.pnas.org/doi/10.1073/pnas.2318106121

## 36. Thermodynamic Machine Learning through Maximum Work Production -

Complexity Sciences Center, https://csc.ucdavis.edu/~cmg/papers/TML.pdf

## 37. Sebastian Goldt - Google Scholar, 

https://scholar.google.it/citations?user=R06wsMkAAAAJ&hl=it

## 38. A deep learning theory for neural networks grounded in physics - Redwood

Center for Theoretical Neuroscience, https://redwood.berkeley.edu/wp-content/uploads/2022/11/scellier-equil-prop.pdf

## 39. Thermodynamic Bound on Energy and Negentropy Costs of Inference in Deep

Neural Networks - arXiv, https://arxiv.org/html/2503.09980v1

## 40. Optimization of Adiabatic Superconducting Logic Cells by Using π Josephson

Junctions, https://www.researchgate.net/publication/374209407_Optimization_of_Adiabatic

_Superconducting_Logic_Cells_by_Using_p_Josephson_Junctions

## 41. Communication consumes 35 times more energy than computation in the human

cortex, but both costs are needed to predict synapse number | PNAS, https://www.pnas.org/doi/10.1073/pnas.2008173118

## 42. Computation in the human cerebral cortex uses less than 0.2 watts yet this great

expense is optimal when considering communication costs | bioRxiv, accessed https://www.biorxiv.org/content/10.1101/2020.04.23.057927.full

## 43. Communication consumes 35 times more energy than computation in the human

cortex, but both costs are needed to predict synapse number - PMC - PubMed Central, https://pmc.ncbi.nlm.nih.gov/articles/PMC8106317/

## 44. The Hidden Cost of AI Energy Consumption - Knowledge at Wharton, accessed

https://knowledge.wharton.upenn.edu/article/the-hidden-cost-of-ai-energy-cons umption/

## 45. Thermodynamic bounds on energy use in Deep Neural ... - arXiv, accessed

 https://arxiv.org/abs/2503.09980

## 46. Heat, Not Halting Problems: Why Thermodynamics May Decide AI Safety -

Medium, https://medium.com/@Elongated_musk/heat-not-halting-problems-why-thermo dynamics-may-decide-ai-safety-6963bfcd3c7c

## 47. Adiabatic Quantum-Flux-Parametron: A Tutorial Review, accessed November 23,

2025, https://search.ieice.org/bin/pdf_advpub.php?category=C&lang=E&fname=2021SE

P0003&abst=

## 48. Margin and Energy Dissipation of Adiabatic Quantum-Flux-Parametron Logic at

Finite Temperature | Request PDF - ResearchGate, https://www.researchgate.net/publication/260515581_Margin_and_Energy_Dissip ation_of_Adiabatic_Quantum-Flux-Parametron_Logic_at_Finite_Temperature

## 49. Beyond Moore's technologies: operation principles of a superconductor

alternative - PMC, https://pmc.ncbi.nlm.nih.gov/articles/PMC5753050/

## 50. Simulation of sub-kBT bit-energy operation of adiabatic

quantum-flux-parametron logic with low bit-error-rate - ResearchGate, accessed https://www.researchgate.net/publication/257955843_Simulation_of_sub-kBT_bitenergy_operation_of_adiabatic_quantum-flux-parametron_logic_with_low_bit-er ror-rate

## 51. Reversibility and energy dissipation in adiabatic superconductor logic - PMC -

NIH, https://pmc.ncbi.nlm.nih.gov/articles/PMC5428326/
52. (PDF) Digital Control of a Superconducting Adiabatic Sigma Neuron -
ResearchGate, https://www.researchgate.net/publication/392996704_Digital_Control_of_a_Super conducting_Adiabatic_Sigma_Neuron
53. (PDF) Learning cell for superconducting neural networks - ResearchGate,
https://www.researchgate.net/publication/347488470_Learning_cell_for_superco nducting_neural_networks

## 54. Adiabatic superconducting cells for ultra-low-power artificial neural networks -

PMC - NIH, https://pmc.ncbi.nlm.nih.gov/articles/PMC5082478/

## 55. Researchers Use AI to Expose Hidden Webs of Entanglement - The Quantum

Insider, https://thequantuminsider.com/2025/09/18/researchers-use-ai-to-expose-hidden
-webs-of-entanglement/