Energentic Intelligence: From Self-Sustaining Systems to Enduring Artificial Life

Atahan Karagoz

REVIEW 5 major objections 5 minor 18 references

Energentic Intelligence: From Self-Sustaining Systems to Enduring Artificial Life

T0 review · 5 major / 5 minor · reviewed 2026-08-07 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Pith's one-line read The paper claims that an agent whose only objective is to keep its cumulative energy surplus non-negative for as long as possible can learn—without external task rewards—to harvest, regulate heat, and compute in a way that keeps it viable.

desk verdict A well-intentioned position paper whose toy demo overclaims: the framework is standard viability theory with a new name, but the research agenda is worth a serious referee. read the letter →

arxiv 2506.04916 v1 pith:NKKR5MZV submitted 2025-06-05 cs.AI cs.LGcs.SYeess.SY

Atahan Karagoz This is my paper

classification cs.AIcs.LGcs.SYeess.SY

keywords EnergenticIntelligenceenergy-basedutilityfunctionsurvivalhorizonself-sustainingagentsresource-awarecomputationthermoregulationviabilitymetricsautonomous

verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

The paper tries to establish that survival—defined as keeping cumulative net energy surplus non-negative for as long as possible—is a sufficient objective for autonomous agents. It formalizes this with an energy-based utility function and a survival horizon, then proposes a modular architecture that couples energy harvesting, computation, and thermal regulation into one control loop. A simulated grid world is used to argue that a Q-learning agent optimized for this survival objective learns to move, idle, and compute in ways that keep it viable under fluctuating energy and temperature, while a fixed-compute policy collapses and a greedy harvester survives without doing anything. If correct, the work would shift AI design from task reward to self-preservation, relevant to robots and devices in settings where power and cooling cannot be assumed.

What carries the argument

The argument is carried by two coupled state-update equations and a single objective. Energy evolves as $e_{t+1} = e_t + \eta P(x_t, y_t) \delta_a - c(a_t)$ and temperature as $T_{t+1} = T_t + \alpha h(a_t) - \beta D(x_t, y_t)$; the survival horizon H is the largest t such that cumulative net energy surplus stays non-negative, and the agent's policy is chosen to maximize H. A Q-learning variant with a viability-tuned reward negotiates between the three actions—move, compute, idle—and the viability metrics EVS, TRI, SHE, and their composite EAS are proposed to score persistence, thermal resilience, and forecasting accuracy.

What would settle it

Re-run the Section V simulation with the agent's observation restricted to its own energy and temperature plus the energy-harvesting and cooling values of its current cell only, removing any global view of P(x,y) and D(x,y); if the survival-optimized policy then no longer moves toward favorable cells but merely idles, the claimed emergence of resource-aware navigation is refuted.

Watch

Extended reading notes

Core claim

On the paper's own terms, the central claim is that an agent whose only objective is to maximize its survival horizon H—the longest time t for which the cumulative energy surplus EUF stays non-negative—will exhibit stable, resource-aware behavior without task-specific rewards. In the simulation, the survival-optimized policy harvests energy where it is available, sheds heat where cooling is available, and modulates computation and dormancy to avoid both energy depletion and thermal runaway. The author presents this as evidence for a new class of 'Energentic Intelligence' in which persistence, not task performance, is the organizing principle of agency.

Load-bearing premise

The load-bearing premise is that the learning agent has enough access to its environment's energy-harvesting and cooling fields that it can learn to move toward favorable cells; the paper never states the observation space, so if that access is missing the reported resource-aware behavior cannot emerge from the given updates alone.

Editorial extensions

If this is right

If survival-horizon maximization is sufficient, autonomous devices can persist in off-grid or post-disaster settings without task-specific reward engineering or stable infrastructure.
The viability metrics EVS, TRI, SHE, and EAS turn persistence into a quantitative design target, allowing systems to be compared by how long and how well they keep themselves viable rather than by task accuracy.
The four-module architecture points to concrete prototypes built from low-power microcontrollers, flexible photovoltaics, thermistors, and capacitive storage, with behavior gated by internal viability thresholds.
The Persistence–Computability Tradeoff question means energy-bounded agents may have to give up arbitrary computation for longevity, making inference depth itself a resource-allocation decision.
Because persistence precedes task completion, an Energentic agent can refuse or defer commands that threaten its viability, which the paper argues calls for coordination and governance protocols rather than simple reward shaping.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

The paper leaves implicit that its survival horizon objective is a no-ruin constraint, so risk-averse and robust decision methods could provably outperform the Q-learning demonstration under uncertain energy forecasts.
A natural next experiment the paper does not report is to hold computation demand fixed and compare survival-trained agents with task-reward agents in the same grid world; if survival-trained agents still perform useful computation, the 'no external supervision' claim becomes directly testable.
The simulation's resource-aware navigation depends on an observability premise: moving toward high-harvest, high-cooling cells requires access to the spatial fields, which the paper never specifies as part of the agent's observation space; if those fields are given as oracle inputs, the demonstration presupposes the predictive ability the limitations section calls impossible.
The paper names but does not model an ecological consequence: in a multi-agent setting, agents prone to extreme conservatism may be out-survived by those taking calibrated risks, suggesting selection pressure could shape the trade-off between engagement and endurance.

Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Desk Editor's Note

read the letter

Quick read for you: this paper is a vision essay dressed as a research paper. The core formal contents—the Energetic Utility Function as an energy balance, the survival horizon H as a cumulative viability constraint—are elementary instances of Aubin's viability theory, which the paper itself cites. If you read it for new results, you'll be disappointed. If you read it as a structured research agenda for energy-viable autonomy, it's actually decent.

What it does well: the four-module architecture is a clean way to organize the discussion, and the EVS/TRI/SHE/EAS metric family gives practitioners a shared vocabulary, even if none of these are derived or validated beyond assertion. The limitations section is candid: it acknowledges forecasting failures, material nonlinearities, and the risk of conservatism-induced stasis. That honesty is rare. The suggested persistence-computability tradeoff is the most interesting question in the paper; it could be a dissertation topic.

Soft spots, in proportion. The central demonstration is the weak link. First, the reward is 'tuned for energetic and thermal viability'—so the agent is directly optimized for what the metrics then measure. The abstract's 'without external supervision' is a stretch. Second, the observation space is never specified. Equations (3) and (4) depend on P(x,y) and D(x,y), so any policy that moves toward favorable cells must observe position or these fields, or a local proxy. The paper's stated internal state s_t=(e_t,T_t,a_t) doesn't include them, and the simulation section never says what the Q-learning agent actually sees. This is load-bearing: either the agent had oracle access, which makes the navigation behavior a consequence of supervision, or it didn't, in which case the reported behavior is unexplained. The paper's own Section VIII says survival-oriented policies depend on forecasts that are 'often inaccurate or impossible,' which sits awkwardly with any oracle assumption. Third, no code, reward values, or hyperparameters are released, so reproduction is blocked. The boundedness assertions for EVS and EAS are also hand-wavy, as the reader's report notes.

None of these are fatal to the paper's value as a position piece, but they are fatal to its claim of a demonstrated emergent behavior. The authors need to reframe the paper as a proposal, disclose the reward and observation vector, and either prove or soften the formal claims.

Who should read it: people working on energy-harvesting embedded systems, budgeted reinforcement learning, and the philosophical side of artificial life. I would not cite the simulation as evidence, but I would bring the paper to a reading group to argue about whether 'viability before performance' is a useful design principle. It deserves a serious referee because the agenda is worth engaging, but I would ask for major revision before publication.

Referee Report

5 major / 5 minor

Summary. The paper introduces "Energentic Intelligence," a framework in which autonomous agents are defined not by task performance but by their capacity to maintain energetic and thermal viability over time. It formalizes an Energetic Utility Function (EUF), a survival horizon H, and a policy objective π* = argmax H, then proposes a modular architecture (Energy Generation Core, Energo-Cognitive Cortex, Thermal Regulation Unit, Survival Manager). The central demonstration is a grid-world simulation in which a Q-learning agent, trained with a custom reward "tuned for energetic and thermal viability," learns to balance harvesting, computation, and idling, outperforming a fixed-compute policy and a greedy harvester. The paper introduces viability metrics (EVS, TRI, SHE) and a composite EAS score, and discusses philosophical and ethical implications of survival-driven agents.

Significance. If made rigorous and reproducible, the paper would offer a useful conceptual contribution to energy-aware and self-sustaining agent design. The idea of treating survival under thermodynamic constraints as the objective, rather than as a constraint on task performance, is worth exploring, and the explicit comparison of fixed-compute, greedy, and survival-optimized policies is instructive. The paper also candidly acknowledges limitations, including forecasting difficulty and behavioral conservatism, which is a strength. However, the formal definitions have mathematical gaps, and the central simulation is not verifiable as written because key details—observation space, reward function, and environment parameters—are omitted. The claim of behavior emerging "without external supervision" is also in tension with the use of a tuned reward function. These issues are load-bearing for the paper's main claims and require substantial revision.

major comments (5)

[Section V, Eqs. (3)-(4) and Section III] The energy and thermal dynamics in Eqs. (3) and (4) depend explicitly on the agent's position (x_t, y_t) and on the spatial fields P(x,y) and D(x,y). However, Section III defines the internal state as s_t = (e_t, T_t, a_t), with no position or field information, and Section V never specifies the observation vector used by Q-learning. Without access to position or to P and D (or a local proxy), the agent cannot learn to move toward energy-rich or cool cells, so the reported "resource-aware" navigation is unexplained. If the agent is given oracle access to P and D, the behavior is presupposed by the environmental model rather than emergent from survival pressure. The paper must specify the observation space, including whether position and field values are observed and at what cost, and should address how this squares with Section VIII's statement that forecasting energy inflow is "often inaccurate or impossible."
[Section V, reward function and Abstract] The custom reward function "tuned for energetic and thermal viability" is never defined. Since Q-learning maximizes this reward, the abstract's claim that the simulation demonstrates "stable, resource-aware behavior without external supervision" cannot be verified: the agent is supervised by a hand-designed reward that explicitly encodes viability. If the reward is essentially a scalarization of the same energetic and thermal quantities measured by EVS, TRI, and EAS, then the high EAS score is largely fixed by construction. The reward function should be stated explicitly, and the sense in which the behavior is "unsupervised" should be clarified. Without this, the central demonstration is not independently checkable.
[Section III, Eq. (2)] The survival horizon H = max_t { t : sum_{τ=0}^{t} EUF_π(τ) ≥ 0 } is not well-defined when the cumulative surplus is nonnegative for all t, since the set over which the maximum is taken is unbounded. Moreover, for a system that terminates when energy is depleted, the relevant quantity is the first time the cumulative surplus becomes negative, not the latest time at which it is nonnegative. As written, H and the resulting objective π* = argmax H are ambiguous, and this ambiguity propagates to SHE in Eq. (7), where H_t is called "the true time until shutdown." The definition should be corrected to a hitting time or an explicit bound should be imposed.
[Section V.A and Section V.D, Eqs. (5) and (8)] EVS is defined as an average of net energy surplus over active periods; unless the environment and action costs are bounded, this quantity is unbounded. The claim that EVS is "bounded and differentiable" (Section V.A) is therefore not true in general, and the same applies to the claim that EAS is bounded (Section V.D), since EAS = EVS * TRI / (1 + SHE) inherits unboundedness from EVS. Differentiability also depends on the smoothing of the indicator and absolute-value terms in Eqs. (5) and (7). The boundedness and differentiability claims should be either proven under explicit assumptions or removed.
[Section V, simulation description] The simulation section reports qualitative results—fixed compute fails by timestep 5, greedy harvesting avoids failure but computes nothing, and the survival-optimized policy maintains a stable trajectory—but does not report the environment generation procedure for P(x,y) and D(x,y), the Q-learning hyperparameters, the number of episodes, the exploration schedule, or the numerical values behind Figures 3-5. No code is released. The claim that a survival-optimized policy "maintains a stable trajectory" is therefore not reproducible. At minimum, the paper should provide the parameter settings and either release the code or report quantitative trajectories and metric values for all three policies.

minor comments (5)

[Figure 5] The text states that the fixed-compute policy "fails by timestep 5," but the horizontal axis in Figure 5 extends to 8; please clarify the time range and mark the failure point explicitly.
[Section V.A, Eq. (5)] The notation E_in(t) and E_out(t) is introduced in Eq. (5), while Eq. (1) uses E_in and E_out without time indices; unify the notation for clarity.
[Section VI] The sentence "This is not a failure of alignment—it is a coherent expression of autonomy under constraint" appears twice in Section VI; remove the duplicate.
[Section IV] The phrase "a closed architecture resembling to a metabolic nervous system" contains a grammatical error; change "resembling to" to "resembling."
[References] Reference [9] is a self-citation to "Computational inertia as a conserved quantity in frictionless and damped learning dynamics" that is invoked several times as an analogy, but the connection to energetic viability is not sufficiently established; either justify the connection in the text or cite it less prominently.

Circularity Check

1 steps flagged · score 6.0 of 10

The claimed 'emergence without external supervision' rests on a hand-tuned viability reward, and the reported EAS measures exactly that reward's target; the central demonstration is therefore not an independent prediction.

fitted input called prediction [Section V (Theoretical Simulation), Eqs. (3)-(8), especially EVS/TRI/EAS definitions and the survival-optimized policy description]
"A Q-learning [11] variant is employed with a custom reward function tuned for energetic and thermal viability. Terminal states are triggered when energy is fully depleted or temperature exceeds a critical threshold. ... Together, these support a high EAS score—indicating not just survival, but well-regulated Energentic behavior."

The Q-learning reward is hand-tuned for 'energetic and thermal viability', and terminal states are exactly energy depletion or exceeding Tcrit. The evaluation metrics EVS (Eq. 5) and TRI (Eq. 6) measure, respectively, average net energy surplus during active intervals and the fraction of time T_t stays below Tcrit; EAS (Eq. 8) is EVS·TRI/(1+SHE). Thus the reported 'high EAS score' is a re-scoring of the same viability objective the reward was built to maximize, not an independent emergent result. Calling this behavior 'without external supervision' additionally overlooks that the custom reward is itself an externally supplied training signal. The claimed prediction is therefore forced by the fitted reward by construction.

full rationale

The central simulated demonstration is not self-contained: the agent's declared internal state s_t=(e_t,T_t,a_t) omits position and the fields P(x,y), D(x,y) that enter Eqs. (3)-(4), so the reported navigation to favorable cells is unexplained unless oracle access is assumed. Independently of that gap, the only quantifiable 'emergence' result—a high EAS for the survival-optimized policy—is aligned by construction with the hand-tuned viability reward used for Q-learning: EVS and TRI measure the same energetic/thermal viability that the reward is tuned to produce, so reporting them as evidence is circular in the sense of a fitted input being called an emergent prediction. The self-citations to the author's prior work [9] are analogical, not load-bearing, and no uniqueness theorem is invoked, so those do not add circularity. However, because the predicted outcome is not derived from first principles and is untestable without the omitted reward specification, observation vector, and code, a moderate circularity score is warranted.

Assumptions & free parameters 7 free parameters · 5 assumptions · 4 invented entities

The framework stands on two kinds of unpaid inputs. First, its defining axioms: the Postulate of Persistence and A1-A3 are asserted in Section III and simply define what counts as an Energentic agent (ad hoc to this paper). Second, the simulation's validity depends on hand-chosen coefficients (eta, delta_a, action costs, alpha, beta, T_crit), an unspecified reward function, unreported Q-learning hyperparameters, and unstated field-generation rules for P and D, together with the linearity assumptions of Eqs. (3)-(4) that the paper's own Limitations section admits are idealizations. The invented entities are conceptual (a paradigm label, a metric family, a module blueprint), each without independent falsifiable leverage. About 7 free parameters plus 5 axioms and 4 invented entities carry the central demonstration.

free parameters (7)

Harvesting efficiency eta and orientation gain delta_a = not stated
Eq. (3) scales ambient potential P into usable energy; values are hand-chosen simulation constants that control whether harvesting admits viable policies in the grid world.
Action costs c(a_t) for move, compute, idle = not stated
Eq. (3) subtracts action cost from energy; these hand-picked values determine which policies are viable and sit at the core of the demonstration.
Heat generation alpha*h(a_t) and dissipation beta*D(x,y) = not stated
Eq. (4) governs thermal dynamics; these coefficients set how quickly compute heats the agent and cooling relaxes it, directly shaping any learned survival policy.
Critical temperature T_crit = not stated
Defines the terminal overheating threshold used in the reward and in TRI (Eq. 6); its value sets the stringency of the thermal constraint.
Custom viability reward function = described only as 'tuned for energetic and thermal viability'
Section V states a custom reward is used; its weights are the actual objective being optimized and are never specified numerically, yet all reported survival behavior depends on them.
Q-learning hyperparameters (step size, discount, exploration, episodes) = not stated
Required to reproduce the learned policy behind Figures 3-5; none are reported.
Environment field generation for P(x,y), D(x,y) = not stated
The spatial energy and cooling landscapes create the fluctuation the agent must survive; generation rules are not given.

assumptions (5)

ad hoc to paper Postulate of Persistence: any agent that does not model its own energetic viability cannot be considered autonomous.
Section III. This definitional postulate is the foundation of the framework; it is asserted, not derived, and it redefines 'autonomous' in a way that makes the paper's thesis true by terminology.
ad hoc to paper A1-A3: no stable external energy source; computation scales with energy and thermal pressure; internal variables are continuously sensed.
Section III lists these as the three operational axioms of Energentic agents; they are assumptions about what counts as an Energentic agent, not consequences of anything prior.
domain assumption Energy and thermal dynamics are linear and Markovian (Eqs. 3-4): harvesting is a reliable signal, cooling is linear feedback.
Section V and Section VIII. The simulation's survival lessons rest on these idealizations; the paper itself flags in Limitations that real harvesting fluctuates chaotically and thermal control has lags and degradation.
domain assumption Cumulative net energy surplus is the correct viability criterion (Eq. 2).
The survival horizon treats viability as non-negativity of the cumulative surplus, ignoring one-period catastrophic deficits or thermal events that do not immediately drive the sum negative.
domain assumption A Q-learning agent with a viability-tuned reward can discover a policy in the described state space.
Section V uses Q-learning without specifying the observation space; learning a navigation-and-dormancy policy requires information about fields P and D that the declared state st=(et,Tt,at) does not contain.

invented entities (4)

'Energentic Intelligence' as a named paradigm
purpose: Package energy-viability-based autonomy as a distinct class of AI with its own objective (persistence) and metrics.
The paradigm is a reframing of existing cybernetic, autopoietic, and viability-theoretic ideas into a new label; it offers no falsifiable handle outside the paper's own simulation.
EVS/TRI/SHE/EAS viability metric family
purpose: Quantify energetic surplus, thermal resilience, forecast error, and composite adaptive stability for Energentic agents.
The metrics are simple statistics over the paper's own state variables; they are calibrated by the authors on their own toy domain with no external benchmark, so scores are not independently anchored.
Four-module architecture (Energy Generation Core, Energo-Cognitive Cortex, Thermal Regulation Unit, Survival Manager)
purpose: Provide an implementable blueprint for survival-driven closed-loop control.
The modules are conceptual proposals; no hardware or firmware implements them, and no component is defined tightly enough to falsify or validate. Section IV.
Eco-machinic life (hybrid bio-electrical agents)
purpose: Speculative future class of agents drawing energy from biological interfaces.
Mentioned in Section VII as a frontier concept, not operationalized; no design or testable prediction is given.

how reviews work

0 comments

Cite this review

Pith. "Pith review of Energentic Intelligence: From Self-Sustaining Systems to Enduring Artificial Life." pith.science (2026). https://pith.science/paper/NKKR5MZV

@misc{pith2026250604916,
  author       = {Pith},
  title        = {Pith review of: Energentic Intelligence: From Self-Sustaining Systems to Enduring Artificial Life},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/NKKR5MZV}},
  note         = {Machine review of arXiv:2506.04916}
}

read the original abstract

This paper introduces Energentic Intelligence, a class of autonomous systems defined not by task performance, but by their capacity to sustain themselves through internal energy regulation. Departing from conventional reward-driven paradigms, these agents treat survival-maintaining functional operation under fluctuating energetic and thermal conditions-as the central objective. We formalize this principle through an energy-based utility function and a viability-constrained survival horizon, and propose a modular architecture that integrates energy harvesting, thermal regulation, and adaptive computation into a closed-loop control system. A simulated environment demonstrates the emergence of stable, resource-aware behavior without external supervision. Together, these contributions provide a theoretical and architectural foundation for deploying autonomous agents in resource-volatile settings where persistence must be self-regulated and infrastructure cannot be assumed.

Figures

Figures reproduced from arXiv: 2506.04916 by the authors.

**Figure 1.** Heatmap showing the agent’s internal state dynamics over time. Each vertical slice corresponds to a single timestep, capturing the simultaneous evolution [PITH_FULL_IMAGE:figures/full_fig_p003_1.png] view at source ↗

**Figure 2.** Subsystem-level architecture of an Energentic agent. Inspired by [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗

Figures from the paper (3 more)

**Figure 3.** Figure 3: Temporal sequence of behavioral state transitions aligned with internal energy levels. The plot tracks how the agent enters dormant, active, or degraded [PITH_FULL_IMAGE:figures/full_fig_p005_3.png]

**Figure 5.** Figure 5: Temporal evolution of internal energy reveals the consequences [PITH_FULL_IMAGE:figures/full_fig_p005_5.png]

**Figure 6.** Figure 6: Viability landscape showing the survival horizon [PITH_FULL_IMAGE:figures/full_fig_p008_6.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

18 extracted references · 13 canonical work pages

[1]

Energy and policy considerations for deep learning in nlp,

E. Strubell, A. Ganesh, and A. McCallum, “Energy and policy considerations for deep learning in nlp,” inProceedings of the 57th Annual Meeting of the Association for Computational Linguistics. Association for Computational Linguistics, 2019, pp. 3645–3650. [Online]. Available: https://aclanthology.org/P19-1355

work page 2019
[2]

On the dangers of stochastic parrots: Can language models be too big?

E. M. Bender, T. Gebru, A. McMillan-Major, and M. Shmitchell, “On the dangers of stochastic parrots: Can language models be too big?” in Proceedings of the 2021 ACM Conference on Fairness, Accountability, and Transparency. Association for Computing Machinery, 2021, pp. 610–623. [Online]. Available: https://doi.org/10.1145/3442188.3445922

arXiv 2021
[3]

Wiener,Cybernetics: Or Control and Communication in the Animal and the Machine

N. Wiener,Cybernetics: Or Control and Communication in the Animal and the Machine. Cambridge, MA: MIT Press, 1948. [Online]. Available: https://mitpress.mit.edu/9780262730099/cybernetics

arXiv 1948
[4]

W. R. Ashby,Design for a Brain: The Origin of Adaptive Behavior. New York, NY: John Wiley & Sons, 1952. [Online]. Available: https://archive.org/details/designforbrain00ashb

work page 1952
[5]

Can deep learning revolutionize mobile sensing?

N. D. Lane, P. Georgiev, and L. Qendro, “Can deep learning revolutionize mobile sensing?” inProceedings of the 16th International Workshop on Mobile Computing Systems and Applications, ser. HotMobile ’15. Association for Computing Machinery, 2015, pp. 117–122. [Online]. Available: https://dl.acm.org/doi/10.1145/2699343.2699349 8

arXiv 2015
[6]

Energy as a constraint on the coding and processing of sensory information,

S. B. Laughlin, “Energy as a constraint on the coding and processing of sensory information,”Current Opinion in Neurobiology, vol. 11, no. 4, pp. 475–480, 2001. [Online]. Available: https: //doi.org/10.1016/S0959-4388(00)00237-3

work page doi:10.1016/s0959-4388(00)00237-3 2001
[7]

Energy scavenging for mobile and wireless electronics,

J. A. Paradiso and T. Starner, “Energy scavenging for mobile and wireless electronics,”IEEE Pervasive Computing, vol. 4, no. 1, pp. 18–27,

work page
[8]

Assessing the potential of artificial intelligence and machine learning for thermal management in electronic devices,

O. Yatskiv and B. Koman, “Assessing the potential of artificial intelligence and machine learning for thermal management in electronic devices,”Technology Audit and Production Reserves, vol. 1, no. 81, pp. 58–74, 2025. [Online]. Available: https: //journals.uran.ua/tarp/article/view/323117

work page 2025

Show all 18 references

[9]

Computational inertia as a conserved quantity in frictionless and damped learning dynamics,

A. Karagoz, “Computational inertia as a conserved quantity in frictionless and damped learning dynamics,”arXiv preprint, 2025. [Online]. Available: https://arxiv.org/abs/2505.19171

2025 arXiv
[10]

Life as a manifestation of the second law of thermodynamics,

E. D. Schneider and J. J. Kay, “Life as a manifestation of the second law of thermodynamics,”Mathematical and Computer Modelling, vol. 19, no. 6-8, pp. 25–48, 1994. [Online]. Available: https://www.sciencedirect.com/science/article/pii/0895717794901880

1994
[11]

Russell and P

S. Russell and P. Norvig,Artificial Intelligence: A Modern Approach, 4th ed. Hoboken, NJ: Pearson, 2020, ch. Chapter 23: Reinforcement Learning, pp. 840–870. [Online]. Available: https://aima.cs.berkeley.edu/

2020
[12]

H. R. Maturana and F. J. Varela,Autopoiesis and Cognition: The Realization of the Living. Dordrecht, Holland: D. Reidel Publishing Company, 1980. [Online]. Available: https://link.springer.com/book/10. 1007/978-94-009-8947-4

1980
[13]

A general model for the origin of allometric scaling laws in biology,

G. B. West, J. H. Brown, and B. J. Enquist, “A general model for the origin of allometric scaling laws in biology,”Science, vol. 276, no. 5309, pp. 122–126, 1997. [Online]. Available: https://www.science.org/doi/10.1126/science.276.5309.122

1997 doi
[14]

Aubin,Viability Theory

J.-P. Aubin,Viability Theory. Boston: Birkh ¨auser, 1991. [Online]. Available: https://link.springer.com/book/10.1007/978-0-8176-4910-4

1991 doi
[15]

Budgeted reinforcement learning in continuous state space,

N. Carrara, E. Leurent, R. Laroche, T. Urvoy, O.-A. Maillard, and O. Pietquin, “Budgeted reinforcement learning in continuous state space,” arXiv preprint, 2019. [Online]. Available: https://arxiv.org/abs/1903.01004

2019 arXiv
[16]

Constrained policy optimization,

J. Achiam, D. Held, A. Tamar, and P. Abbeel, “Constrained policy optimization,” inProceedings of the 34th International Conference on Machine Learning (ICML). PMLR, 2017, pp. 22–31. [Online]. Available: http://proceedings.mlr.press/v70/achiam17a.html

2017
[17]

Microbial fuel cells: methodology and technology,

B. E. Logan, B. Hamelers, R. Rozendal, U. Schr ¨oder, J. Keller, S. Freguia, P. Aelterman, W. Verstraete, and K. Rabaey, “Microbial fuel cells: methodology and technology,”Environmental Science & Technology, vol. 40, no. 17, pp. 5181–5192, 2006. [Online]. Available: https://pu...

2006 doi
[2005]

Available: https://ieeexplore.ieee.org/document/1401839

[Online]. Available: https://ieeexplore.ieee.org/document/1401839

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.

[1] [1]

Energy and policy considerations for deep learning in nlp,

E. Strubell, A. Ganesh, and A. McCallum, “Energy and policy considerations for deep learning in nlp,” inProceedings of the 57th Annual Meeting of the Association for Computational Linguistics. Association for Computational Linguistics, 2019, pp. 3645–3650. [Online]. Available: https://aclanthology.org/P19-1355

work page 2019

[2] [2]

On the dangers of stochastic parrots: Can language models be too big?

E. M. Bender, T. Gebru, A. McMillan-Major, and M. Shmitchell, “On the dangers of stochastic parrots: Can language models be too big?” in Proceedings of the 2021 ACM Conference on Fairness, Accountability, and Transparency. Association for Computing Machinery, 2021, pp. 610–623. [Online]. Available: https://doi.org/10.1145/3442188.3445922

arXiv 2021

[3] [3]

Wiener,Cybernetics: Or Control and Communication in the Animal and the Machine

N. Wiener,Cybernetics: Or Control and Communication in the Animal and the Machine. Cambridge, MA: MIT Press, 1948. [Online]. Available: https://mitpress.mit.edu/9780262730099/cybernetics

arXiv 1948

[4] [4]

W. R. Ashby,Design for a Brain: The Origin of Adaptive Behavior. New York, NY: John Wiley & Sons, 1952. [Online]. Available: https://archive.org/details/designforbrain00ashb

work page 1952

[5] [5]

Can deep learning revolutionize mobile sensing?

N. D. Lane, P. Georgiev, and L. Qendro, “Can deep learning revolutionize mobile sensing?” inProceedings of the 16th International Workshop on Mobile Computing Systems and Applications, ser. HotMobile ’15. Association for Computing Machinery, 2015, pp. 117–122. [Online]. Available: https://dl.acm.org/doi/10.1145/2699343.2699349 8

arXiv 2015

[6] [6]

Energy as a constraint on the coding and processing of sensory information,

S. B. Laughlin, “Energy as a constraint on the coding and processing of sensory information,”Current Opinion in Neurobiology, vol. 11, no. 4, pp. 475–480, 2001. [Online]. Available: https: //doi.org/10.1016/S0959-4388(00)00237-3

work page doi:10.1016/s0959-4388(00)00237-3 2001

[7] [7]

Energy scavenging for mobile and wireless electronics,

J. A. Paradiso and T. Starner, “Energy scavenging for mobile and wireless electronics,”IEEE Pervasive Computing, vol. 4, no. 1, pp. 18–27,

work page

[8] [8]

Assessing the potential of artificial intelligence and machine learning for thermal management in electronic devices,

O. Yatskiv and B. Koman, “Assessing the potential of artificial intelligence and machine learning for thermal management in electronic devices,”Technology Audit and Production Reserves, vol. 1, no. 81, pp. 58–74, 2025. [Online]. Available: https: //journals.uran.ua/tarp/article/view/323117

work page 2025