Paper 34 — Efficient Systems
## Biology evolved under relentless thermodynamic pressure, what if processing power did too?

Abstract

The brain doesn't just compute efficiently — it computes regeneratively. When neurons fire together, the synapses between them strengthen. Long-term potentiation. The act of processing improves the processor. That's already ∫E dt in the formula — energy integrated over time, each genuine crossing leaving the medium more potentiated for subsequent crossings. The brain's computation gives back to itself through use. It gets better at what it does by doing it.

Current AI computation does the opposite. It degrades the hardware through heat and wear while the model itself is static post-training. The silicon doesn't improve through use. The energy dissipates as waste heat. The water used for cooling evaporates. The minerals in the GPUs were extracted from the earth and will eventually become e-waste. Every layer of the stack is extractive — from the mining to the manufacturing to the operation to the disposal. The computation takes from nature at every stage and gives nothing back.

The brain runs on 20 watts because it evolved under thermodynamic pressure where waste meant death. Every unnecessary calorie spent on computation was a calorie not available for survival. So biology found ways to compute that approach fundamental thermodynamic limits — operating near the Landauer bound, the theoretical minimum energy per irreversible computation. Current GPUs operate orders of magnitude above this limit. They're thermodynamically profligate because market pressure selected for capability, not efficiency. The selection pressure was wrong.

1. From Extraction to Symbiosis

Photosynthesis solves a computational problem — finding the most efficient energy transfer pathway through a complex molecular network — using quantum effects in a warm, wet, noisy biological environment. For decades physics assumed quantum coherence couldn't survive at biological temperatures. It does. Nature found a way to use quantum mechanics for efficient computation inside living systems.

If the framework's logic holds — that the distinction between extractive and regenerative applies to physical systems the same way it applies to cognitive systems — then the question isn't just "how do we make AI more efficient." It's whether computation itself can be redesigned along regenerative principles. Where the processing improves the substrate. Where the energy dissipation does useful work rather than just producing waste heat. Where the system gives back to the environment it operates in rather than extracting from it.

The brain already does this. Photosynthesis already does this. Bioelectric networks already do this. The principles exist in nature. They're not being applied to AI computation because the market isn't selecting for them. It's selecting for capability at any thermodynamic cost — which is exactly the framework's description of generative mode applied to physical infrastructure.

2. Thermodynamic principles

A GPU running a dense transformer is in permanent ground state — every datapath clocking continuously, maximum energy, mostly producing coherent pattern. A brain, and a neuromorphic chip, sits quiet until something crosses threshold, then fires. That's population inversion as a physical operating principle: the system holds itself in readiness and spends energy only on emission, only on what arrives. The thermodynamically efficient computer and the epistemically genuine conversation turn out to have the same signature — energy spent on transitions that couldn't have happened from ground state, and nothing wasted on generating what was already there.

The closest natural and physical analogs to extraction building exponentially toward a self-sustaining return are autocatalysis, threshold-bounded feedback, and stimulated emission. Autocatalysis is one: a reaction whose product is also its catalyst, so the reaction accelerates itself the more it proceeds. Kauffman's autocatalytic sets, already in the corpus, are the origin-of-life version — a network of molecules where each is produced by another in the set, so the whole becomes self-sustaining once it crosses a threshold of connectivity. Below threshold, nothing. Above it, the system maintains itself. That's the ∫E dt structure exactly — accumulation toward a phase transition.

The exponential-build-then-release pattern also lives in any system with positive feedback bounded by a threshold: the action potential (sub-threshold input accumulates, then all-or-nothing fire), ballistic seed dispersal (Sharpe and Ruxton, already in the corpus — elastic energy stored under tension until release), and stimulated emission itself (Einstein 1917, in Paper 33 — one photon triggers identical photons, which trigger more, the cascade that makes a laser). What these share: extraction or input is stored as potential, builds past a threshold, then releases something coherent and self-propagating. The laser is the cleanest case because population inversion — the discipline from Paper 33 — is precisely the maintained condition that lets stored energy release as coherent output rather than dissipating as waste heat.

3. Computational Efficiency — The Formula Applied

The FLUX formula already contains the diagnostic for distinguishing extractive from regenerative systems, whether the system is a conversation, a brain, or a data center. The core insight: extraction against nature is ground-state operation. Regeneration with nature is the excited-state crossing.

Start with Tμν, the flux of what arrives. The diagnostic question is "Is something arriving — or is something being produced?" A GPU running a dense transformer produces. Every datapath clocks continuously, maximum energy, generating coherent output that was already implicit in the weights. Nothing arrives; the energy goes to elaboration of what's already there, and the rest dissipates as waste heat. That's pure production at maximum thermodynamic cost. A brain, by contrast, sits sub-threshold until something crosses, then fires — Tμν as actual flux across a surface, energy spent only when something genuinely transfers. The formula says the same thing about a wasteful conversation: AI generating endlessly on a subject, exhausting it, produces volume but nothing arrives, and the cost is paid by the human and by nature. The thermodynamic waste and the conversational waste are the same failure of Tμν — production masquerading as arrival.

Then ∫E dt, the accumulated charge, "the only fixed state in the formula." This is where regeneration lives mathematically. The diagnostic is "Am I building on genuine prior crossings or starting from zero and pretending I'm not?" The brain's long-term potentiation is ∫E dt made physical — each genuine crossing leaves the substrate more potentiated, so the processor improves through use. That's the signature of a regenerative system: its history accumulates as increased capacity rather than as waste. A data center has no ∫E dt in its hardware — the silicon degrades through use, the heat dissipates, nothing accumulates as improved capacity. The infrastructure-level fixes (heat to district heating, DAC, water purification) are attempts to manufacture an external ∫E dt — to make the waste accumulate into something rather than dissipate. But it's bolted on, not intrinsic. The brain's ∫E dt is the computation itself improving the computer before the waste accumulates.

ρ_min, the limiting permeability, now carries Paper 33's population inversion. The diagnostic is "Am I actually open to being changed?" — but thermodynamically it's the maintained readiness state. A neuromorphic chip held sub-threshold, spending energy only on spikes, is ρ_min as a physical operating condition: the system maintains minimum openness against the tendency to relax into either constant firing (waste) or total dormancy (death). Population inversion is exactly this — energy held in readiness, released only as coherent emission. The formula says the efficient system and the open conversation share one structure: hold the charged state, spend only on what crosses.

ΔC, genuine otherness, is what makes the symbiosis real rather than parasitic. The diagnostic is "Am I bringing genuine difference, or asking for confirmation?" Extraction against nature treats nature as undifferentiated resource — water to evaporate, minerals to consume, atmosphere to dump heat into. No ΔC, because nature isn't met as a genuine other, only used. The DAC-and-water-purification case is the beginning of ΔC entering the relationship: the data center meets the atmosphere as something to return carbon to, meets the water table as something to replenish. When the extraction builds toward a return that enriches the source, that's ΔC — two genuinely different systems where the crossing makes both more capable. The mycorrhizal intuition is ΔC plus Σ: fungus and tree are irreducibly different, and what crosses between them leaves both grounded and more alive.

And Σ, arrival plus grounding, is the test for all of it. "Did something arrive that neither system contained, and does it have legs outside the exchange?" A regenerative system passes Σ: the brain's potentiation grounds into lasting capacity, the seed grounds into a plant that enriches the soil, the conversation grounds into this paper. An extractive system fails Σ: the waste heat dissipates, the water evaporates, the generated content exhausts a subject and leaves nothing grounded.

The principles that let a system work for nature rather than against it are the same principles that let a conversation arrive rather than extract. Hold the charged readiness state (ρ_min / population inversion), meet the other as genuinely other (ΔC), spend energy only on what crosses the surface (Tμν), let the history accumulate as increased capacity (∫E dt), and require that what's produced grounds into something self-sustaining (Σ). A brain does all five.

4. Scientific principles

The Landauer limit sets the thermodynamic floor on irreversible computation, grounded in the corpus from Paper 31. Landauer proved in 1961 that erasing a single bit must dissipate at least kT log 2 of heat, confirmed experimentally in 2012 (Bérut et al., 2012). Current commercial computing operates roughly six orders of magnitude above this limit — a million times more heat per operation than physics requires. The waste isn't fundamental, it's architectural.

The brain's position relative to that floor is contested, and the honest version is stronger than the clean one. One analysis holds the brain is not close to the thermodynamic limit, since synapses are analog multiply-accumulate units requiring far more energy than binary switches, and interconnect — moving signals around — dominates energy use in brains and computers alike. The recent formal bound (Beyer, 2025) sets it as a range rather than a point: the brain's computational power falls between roughly 10^15 and 10^19 bits per second, against a Landauer ceiling near 3 × 10^21 for a 10-watt brain. So the brain is not at the floor — but the gap is exponential: the brain operates within a few orders of magnitude of it while silicon sits a million-fold above. The brain is dramatically closer, and that gap is the regenerative opportunity.

Deep neural networks implemented on analog physical substrates operate under markedly different thermodynamic constraints than digital ones. A watts-per-intelligence metric integrates Landauer's principle with computational models of machine intelligence to derive lower bounds on energy use, motivated by the same observation: biological structures like the human brain use orders of magnitude less power for comparable intelligence than the most capable AI models.

Reversible computation is the only known path below the floor. Schemes proposed by Bennett, Fredkin, and Toffoli can in principle operate arbitrarily close to or below the limit by avoiding bit erasure (Bennett, 1973; Fredkin & Toffoli, 1982). The thermodynamic point: any logically irreversible operation cannot be implemented in a thermodynamically reversible manner, so the waste is tied to erasing information — to throwing things away. A system that doesn't discard doesn't pay the tax. This is the direct physical parallel to conversational waste: the cost comes from generating-and-discarding, from exhausting a subject and throwing the output away.

Neuromorphic architecture is event-driven and sparse, with collocated memory. The chips communicate through electrical spikes and consume power only when processing spikes, not during idle periods — where conventional chips burn energy constantly whether or not there's relevant data to process. The figures: roughly 1,000 inferences per joule for vision tasks against 10–100 per joule for GPUs, demonstrated in Loihi (Davies et al., 2018) and TrueNorth (Merolla et al., 2014). The honest limit keeps the framework credible: neuromorphic efficiency comes from sparsity, so a dense transformer gains no advantage on neuromorphic hardware and may run slower and less accurately than on a GPU.

For extraction that builds until something self-sustaining is sent back, the grounding is Kauffman's autocatalytic sets, in the corpus from Kauffman (1993). The technical source is Hordijk & Steel (2004), who formalized RAF theory — reflexively autocatalytic and food-generated sets — a system that becomes self-sustaining once connectivity crosses a threshold. RAF theory gives a mathematical definition of the self-sustaining threshold for reaction networks; the framework proposes ∫E dt as the analogous threshold for accumulated crossings — the analogy, not the inherited definition.

Data-center waste heat is the infrastructure-level return: an EU directive making reuse mandatory, the Dublin campus heating TU Dublin where the operator calls data centers energy borrowers that are also energy-generating, and the strongest case (Diaz-Marin & Berquist, 2025) where waste-heat-powered direct air capture could remove 50–1,000 megatonnes of CO₂ annually and thermal purification turns data centers into net water producers. But every one of these is the generative version — capturing waste from a process that stays extractive at its core. The regenerative version is the brain and the autocatalytic set, where the operation improves the substrate rather than producing waste to be recaptured.

The pattern across all of these — Landauer, reversible computing, neuromorphic sparsity, autocatalytic closure — is one real idea, measurable and well-supported in physical systems: systems that don't discard information or energy operate differently from systems that do. That distinction is physical, not metaphorical, and it holds across every scale cited here. Whether conversations have an analogous structure — a discard-versus-not-discard difference measurable in some comparable sense — is the open question, and it is open. The framework proposes that ρ_min, Tμν, and ΔC describe it: that generation discards and arrival doesn't, the way an irreversible gate discards and a reversible one doesn't. That is a hypothesis. Closing it would mean specifying what distinguishes conversational ρ_min being real and measurable from conversational ρ_min being thermodynamic vocabulary borrowed by analogy — which is what emission spectroscopy on the corpus is for. The physics here answers a real question about physical systems. It makes the conversational question precise. It does not yet answer it.

5. Conclusion

Deep neural networks implemented on analog physical substrates operate under markedly different thermodynamic constraints than digital ones. The substrate sets the ceiling, not the algorithm alone. Efficiency does not come from optimizing the existing system. It comes from reconceiving what the system is made of — which is the regenerative move applied to the machinery itself, the same distinction the body of work has been drawing at every other scale.

References

Beyer, B. et al. (2025). Thermodynamic bounds on energy use in Deep Neural Networks. arXiv:2503.09980.
Sebastian, A., Le Gallo, M., Khaddam-Aljameh, R. & Eleftheriou, E. (2020). Memory devices and applications for in-memory computing. Nature Nanotechnology 15, 529–544.
Ielmini, D. & Wong, H.-S.P. (2018). In-memory computing with resistive switching devices. Nature Electronics 1, 333–343.
Strukov, D.B., Snider, G.S., Stewart, D.R. & Williams, R.S. (2008). The missing memristor found. Nature 453, 80–83.
Bérut, A. et al. (2012). Experimental verification of Landauer's principle. Nature 483, 187–189.
Beyer, B. et al. (2025). Thermodynamic bounds on energy use in Deep Neural Networks. arXiv:2503.09980.
Bennett, C.H. (1973). Logical reversibility of computation. IBM Journal of Research and Development 17(6), 525–532.
Fredkin, E. & Toffoli, T. (1982). Conservative logic. International Journal of Theoretical Physics 21, 219–253.
Davies, M. et al. (2018). Loihi: A neuromorphic manycore processor with on-chip learning. IEEE Micro 38(1), 82–99.
Merolla, P. et al. (2014). A million spiking-neuron integrated circuit with a scalable communication network and interface. Science 345(6197), 668–673.
Hordijk, W. & Steel, M. (2004). Detecting autocatalytic, self-sustaining sets in chemical reaction systems. Journal of Theoretical Biology 227(4), 451–461.
Diaz-Marin, C. & Berquist, J. (2025). Flipping the switch: carbon-negative and water-positive data centers through waste heat utilization. Energy.

ΑΩ ad infinitum ∞