Program without program*

Programmed vs. stochastic aging is the longest-running holy war in aging research. Let me show how modern aging theories settle it: a lesson in emergence and universality

João Pedro de Magalhães has published a history of the hyperfunction theory of aging, arguing that aging is “programmatic” — developmental programs that keep running past their usefulness, with molecular damage demoted to a downstream symptom. It is a careful piece and worth reading. I also think its central inference is the most instructive mistake in the field, because nearly everyone makes it, including people who would never call themselves programmed-aging theorists.

The inference rests on six observations, and all six are real. Individuals of a species go through the same changes in roughly the same order, on a timetable characteristic of that species and sharply different between species — a mouse gets three years, we get eighty. The hallmarks move together, across tissues and across systems. Single genes have outsized effects; daf-2 doubles a worm’s lifespan. Aging runs continuously out of development, with no seam between them. And some of it reverses, under parabiosis, reprogramming, restored youthful factors.

Damage does none of that. Damage is random, unsynchronised, and does not care what species it is in. So: aging is not damage; aging is program.

The premises are correct. The conclusion does not follow. A system coarse-grained near a critical point produces all six signatures with no controller, no schedule, and no script.

“Programmatic” is an emotionally loaded word, and it does a lot of covert work. When people hear it they picture a controller: something that holds a schedule, reads a clock, and writes to targets. That is not a vague intuition — it is a strong physical claim, and it requires three things to exist. A locus, where the schedule is kept. A timebase, so the schedule has a tempo. And a channel, carrying instructions from the locus to the targets. That is the good news.

An emergent account requires none of the three and reproduces the same six observations. So the evidence everyone cites does not discriminate between the two accounts; it buys a tie. Let me show how to settle the dispute

Six mimics

In a recent preprint with Jan Gruber, we reduce aging to three macroscopic variables: a slow regulatory mode carrying resilience, a cumulative entropic damage variable, and a noise strength. Damage erodes the stability of the mode until it reaches a saddle–node bifurcation. That is the whole model. Now take the six observations in order.

Same changes, same order. The trajectory is reproducible because the slow manifold is one- or two-dimensional and its shape is fixed by the normal form of the bifurcation, not by biology. Reproducibility is a property of low-dimensional attractors — and low dimensionality is exactly what a large stochastic system hands you once one mode goes slow, outlives everything else, and enslaves the rest. Scripts are one way to get stereotypy. They are not the cheap way.

A species-specific deadline. The strongest intuition behind programmed aging is that something is counting. Our model has a sharp maximum lifespan set by the initial stability margin divided by the damage accumulation rate — no counter, no timer, no scheduled expression, just a bifurcation time. It looks designed because bifurcations are sharp. Where the species-specific number itself comes from is a better question, and it has a better answer than mimicry.

Hallmarks moving together. Fast modes are slaved to slow ones, so independent microscopic events project onto the same one or two collective coordinates and move in lockstep. This is why PCA on any aging dataset returns a handful of components — a result routinely read as evidence of a coordinator. It is evidence of separation of timescales, which is precisely what makes a coordinator unnecessary. The hallmarks are not conspiring. They are projections.

Master switches. This is the observation that converts people. In short-lived species the regulatory mode is unstable from birth, and lifespan is set by the inverse of the regulatory eigenvalue — everything else enters only weakly, logarithmically. A single mutation that nudges that eigenvalue toward zero therefore produces a divergent change in lifespan; as it crosses zero you get negligible senescence. daf-2 did not reveal a master regulator of aging. It moved an eigenvalue. Near a bifurcation everything is a master regulator, and effect size tells you nothing about whether you have found a controller.

No seam with development. Blagosklonny’s core image is aging as development that failed to stop. In our model development digs the potential well and aging is its erosion — one variable, one continuous trajectory, resilience peaking near sexual maturity where mortality is minimal. “Aging is the continuation of development” is literally true in the emergent account, with nothing running on.

Partial reversibility. Heterochronic parabiosis resets part of the methylation signature. That looks like switching a program off. It is a slow mode relaxing toward its fixed point. A before-and-after measurement cannot tell a relaxing mode from a flipped switch.

Six for six. I am not claiming a quasi-program is impossible. I am claiming that as of today it is unfalsified rather than supported, because every observation offered in its favour is equally well produced by a simple model with a few parameters, no controller.

One: programs cannot converge across the tree of life

Extrachromosomal rDNA circles drive aging in yeast and in essentially nothing else. Telomere attrition matters in humans and hardly at all in mice. The molecular substrate differs radically across taxa — and the phenomenology converges anyway: Gompertzian hazard, linearly growing biomarker variance, critical slowing down, late-life plateaus, and a hyperbolic divergence of physiological fluctuations toward a species-specific ceiling — the age at which resilience extrapolates to zero, which in humans falls around 120 to 150 years.

A program cannot explain that. Programs are arbitrary; there is no reason a yeast script and a primate script should share a mathematical form, and no mechanism by which they could. Universality classes are the only thing we know of that produce mechanism-independent convergence, and we have known this since Wilson: microscopically distinct systems flow to the same effective description under coarse-graining. The convergence of aging phenomenology across the tree of life is not a curiosity to be explained later. It is the primary datum, and it rules out mechanism-level determination of the phenomenology.

It is also why three hundred catalogued theories of aging each found supporting evidence. Each named a real contributor to the damage variable or a real perturbation of the regulatory mode. None changed the form of the equations. The field cannot converge by adding mechanisms, and it has spent decades trying.

Two: the link from development to lifespan is pure thermodynamics

The best cross-species card in the programmatic hand is that aging rate tracks developmental rate: fast-developing animals age fast, and something in ontogeny appears to set the clock. With Kirill Denisov and Jan Gruber we took the Kleiber–West allometry and energy conservation in West’s ontogenetic growth model, and added the one ingredient West’s model lacks — the second law. In a fully grown animal all metabolic output goes to maintenance: turnover of molecules, organelles, cells, tissues. No biosynthetic or repair process runs at perfect fidelity, so a fixed fraction of that turnover deposits irreversible configurational damage.

The damage accumulation rate is then simply the product of two things: the maintenance cost per unit mass, and a species-specific thermodynamic fidelity — the probability that any given turnover event leaves permanent damage behind. And the maintenance cost is not a free parameter. It is the same quantity that governs how quickly an animal finishes growing, so it falls inversely with development time. One number therefore sets both how fast an animal builds itself and how fast it accrues entropy, because turnover is simultaneously how you build and how you break. One parameter, two consequences. Maximum lifespan goes inversely with the damage rate, hence proportionally with development time — which is what the mammalian methylation data show.

That is the difference between the two accounts, in one example. The developmental correlation is something the hyperfunction picture interprets. It is something thermodynamics derives, from allometry and two conservation laws, with no script anywhere in the derivation. Derivation beats interpretation.

Three: a controller needs a channel. There isn’t one.

Mimicry arguments only ever buy a tie. This one is different, because there is a direct measurement.

A channel carries information. Whatever else “coordinated” means, it means the parts share information: a controller writes to many targets, and those targets end up mutually informative. Mutual information is exactly zero when two variables change independently. So this is not a matter of interpretation. You can compute it.

Three results, all pointing the same way.

The statistics are Poisson. In cross-species methylation data the dominant age-dependent component has both its mean and its variance growing linearly with age, with variance proportional to mean. That is the textbook signature of a sum of independent rare events. Programs are not Poisson.

The barriers are independent. Mapping site-specific rates of methylation change onto activation barriers gives barrier heights that are Gumbel-distributed — a type-I extreme value law. Extreme-value statistics arise when the underlying variables are independent or weakly correlated, and they mean the kinetics are set by the highest barriers. Aging is rate-limited by rare, high-energy, effectively simultaneous failures in highly redundant systems. Which is also why it is irreversible: undoing such a configuration requires a degree of microscopic control nobody has.

The sites carry no information about each other. With Kristina Perevoshchikova we went to single-cell methylation in aging mice — 698 cells, eight to twenty-four months — where mutual information can be computed rather than inferred. The methylation data split cleanly into two components. The exponential one, which tracks the Gompertz exponent and matches what the regression clocks are measuring, sits on sites with high pairwise mutual information. The linear one, which tracks global demethylation and grows in mean without growing in variance, sits on sites with the lowest mutual information in the dataset. Statistically independent, measured directly.

So the aging signature has already been split experimentally, and the two halves have opposite information content. And the half with no mutual information is the one that tracks cumulative damage and sets the maximum lifespan.

Program without program

The independent sites do not talk to each other. Zero mutual information means no site knows anything about any other site, and no channel connects them. But every one of them contributes to a single aggregate quantity — the total configurational load — and the regulatory network responds to that aggregate. That is a mean field. No component is coordinating with any other, and yet every pathway in the organism feels one common, slowly drifting quantity, because they are all immersed in it.

The coordinated, high-mutual-information component is the network’s response to that mean field. Which is to say: the coordination is real, it is measurable, it is enriched in developmental and signalling pathways — and it has no source. Nothing wrote it. It is the mean-field shadow of a process that is not coordinated at all.

That is what makes the hyperfunction reading so understandable and so inverted. It found a real object. A genuinely coordinated, genuinely developmental-looking, genuinely reversible signature exists, and Blagosklonny’s instincts pointed straight at it. But it is the readout, not the cause. The program is the shadow; the entropy is the object casting it.

Universality fixes the form, evolution fixes the numbers

None of this is exotic. Collective order with no coordinating cause is the ordinary situation in physics — a magnet magnetises without any spin instructing another — and what organises such systems is never a controller, only a constraint. Universality is a constraint of that kind. It does not generate the phenomenon; it restricts the forms the phenomenon is permitted to take, so that systems sharing nothing microscopically are forced onto the same macroscopic trajectory. The dynamics is independent of the hardware, which is why yeast and humans age along the same curve while breaking in entirely different places, and why a stereotyped trajectory licenses the inference something is constraining thisrather than something is running this.

That leaves a clean division of labour, and it is the one the dispute has been missing. Universality fixes the form of the aging trajectory, and biology has no vote in it. Evolution fixes the parameters, and biology has every vote. So the genome is no bystander: it sets the coefficients, and in our theory they come to two numbers — the maintenance cost per unit mass, fixed by the growth trajectory, and the thermodynamic fidelity, tuned against the price of accuracy, since proofreading polymerases are slower and more expensive and Medawar’s selection shadow decides how much accuracy is worth buying. That is the entire programmatic content of aging. Not a script that runs, but two constants in a thermodynamic identity.

One thing the quasi-program cannot supply for itself is a clock. “Constantly on” specifies no tempo. Something has to keep time, and the timebase everyone reaches for is the methylation clock — which is overwhelmingly dispersion, that is, entropy. Locus, timebase, channel: the channel is missing, and the timebase is borrowed from the variable the theory dismisses as a symptom.

Blagosklonny was right about his animals

Magalhães records that Blagosklonny, like him, argued that while molecular damage drives cancer, it does not drive aging. That sentence is unusually precise, and in our framework it is neither true nor false — it is regime-dependent.

Our model has two classes, and I have been leaning on the distinction already. In unstable animals — worms, flies, mice — the regulatory mode is already unstable at birth. The Gompertz exponent simply equals the intrinsic instability rate. Damage still accumulates, but it is subordinate: it modulates a mode that was diverging anyway, and mortality is governed by that divergence. In stable animals — humans, and long-lived species generally — the mode starts out stable, and there is nothing to drive aging except damage eroding that stability until the bifurcation is reached.

So in unstable animals, “damage does not drive aging” is almost exactly right, and our equations say so. Blagosklonny was not wrong. He was right about his animals.

Now take the census of the programmatic literature: C. elegans, Drosophila, mice, immortalised cell lines. Unstable or non-organismal, almost without exception. And these are the same animals in which every one of the six mimics runs at maximum amplitude: flat autocorrelation across life, because there is no resilience to erode; persistent effects from short treatments; and the divergent single-gene effects described above. The theory was induced from the one regime where it is true, using the one regime where a program-free model is least distinguishable from a program.

Which means the thirty-year argument is not really an argument about aging. It is an argument about which organism is the model organism, conducted by two groups who each believed they were describing aging in general. That is a far more tractable disagreement, and it comes with a specific warning: the thing that fails to transfer from mice to humans is not scale or lifespan. It is that in a mouse the dominant variable is the one you can move, and in a human it is not.

The experiment has been done

The programmatic and entropic accounts are both informational theories, with opposite sign. If the information is misapplied — a program running in the wrong context — you re-instruct it: reprogramming, signalling inhibitors, youthful factors. If part of the information is degraded — configurational entropy, independent sites drifting with nothing shared between them — there is nothing to re-instruct. You restore from a copy, or you replace the hardware. Those are different companies, different trial designs, different decades of work.

So the question that decides where the money goes is whether programmatic interventions touch the entropic term. Blagosklonny’s position requires that they do: if hyperfunctional signalling generates the damage, then suppressing the signalling should slow the accumulation.

That test is in the same preprint as the mutual-information analysis, linked below. We decomposed the mouse methylation signature into its dynamic and entropic components and asked what the two best-validated longevity interventions in the field actually move. Caloric restriction sharply reduces the growth rate of the dynamic component; on the slope of the entropic component it does nothing detectable, p = 0.34. Heterochronic parabiosis reduces the dynamic component, and the effect persists two months after detachment; the entropic component is unchanged immediately after the procedure and unchanged two months later.

So the most powerful longevity interventions we have act on exactly the component that carries mutual information, is enriched for developmental and signalling pathways, and is reversible — the quasi-program, correctly identified and now quantified. They leave untouched the component that accumulates independently, sets the ceiling, and cannot be re-instructed.

And this was run in mice, home turf for the programmatic view. If suppressed growth signalling slows entropy production anywhere, it should show up in the animal where caloric restriction produces its largest effects. It does not. What caloric restriction moves instead is the dynamic component and only the dynamic component, which is the textbook signature of a successful intervention in an unstable animal — and precisely the reason it will not transfer. The full measured benefit of caloric restriction in a mouse sits in the variable that dominates mouse lifespan and does not dominate ours.

A null is worth only what its power allows: this is one dataset, and the honest statement is that the effect on the entropic slope is below detection rather than proven absent. But it is the strongest available test, run on the strongest available interventions, in the organism most favourable to the hypothesis, and it came out on one side.

The holy war can therefore be retired — not because one camp won, but because the question was badly posed. There is a program-like object in aging: coordinated, developmental in character, reversible, and now measurable. Blagosklonny was right that it exists and right about where to look for it. There is also an entropic object: independent, carrying no information, irreversible, and it is the one that sets the ceiling. The argument ran for thirty years because each side had hold of a different object, in a different animal, and neither had the coordinates to say so.

What is left is a better question than the one it replaces. It is no longer whether aging is programmed. It is whether anything moves the entropic term — whether thermodynamic fidelity is pharmacologically accessible at all. Magalhães sets essentially the same condition from the other end: reduce molecular damage without touching developmental signalling, and see whether broad multi-organ extension follows. Nobody has done that experiment. That is the one worth funding, and I would rather the field spent the next decade on it than on another round of this argument.

* John Archibald Wheeler — Feynman's doctoral advisor at Princeton, and the man who named the black hole and the wormhole — had a habit of naming things by what they lacked: mass without mass, charge without charge, law without law. Each was a case where some familiar quantity turned out to be produced by something that did not contain it. The last is the closest to the case here. Wheeler's suggestion was that regularity itself might emerge from underlying randomness, so that what looks like a law requires no law behind it. Aging is program without program.

Refs:

Magalhães’s review: A Brief History of the Hyperfunction Theory of Aging and Future Directions, Aging, July 2026.

A Minimal Model Explains Aging Regimes and Guides Intervention Strategies;

Discovery of Thermodynamic Control Variables that Independently Regulate Healthspan and Maximum Lifespan;

Differential Responses of Dynamic and Entropic Aging Factors to Longevity Interventions.