The shape without changing the parts
A single genotype can build more than one body. That is the whole of it, stated plainly, before any qualification. The DNA sequence stays fixed; what the organism becomes does not. A water flea exposed to chemical traces of a predator grows a helmet it would never grow in clean water. A seedling raised in the shade of its neighbours stretches its stem and thins its leaves, chasing light it has not yet lost. Nothing in the genome altered between the two outcomes. What altered was which parts of the genome got switched on, in what sequence, in response to what was sensed.
Biologists call the underlying rule a reaction norm: a mapping from environmental input to realised form. The same genotype, run against different conditions, traces out different phenotypes, and the mapping itself is a property that can be measured, compared across lineages, and — this is the detail that keeps the concept from being a curiosity — inherited and evolved in its own right. Natural selection does not only act on which traits an organism has. It acts on how responsive the organism is permitted to be. Some lineages evolve broad, flexible reaction norms. Others evolve narrow ones, or none, committing to a single developmental path regardless of what the environment does. Both are outcomes of selection. Neither is the default.
This matters because it separates two things that look similar from a distance but are not. Genetic adaptation requires differential survival: some genotypes reproduce more than others, and the population's genetic composition shifts across generations. Phenotypic plasticity requires nothing of the sort. It happens inside one organism, inside one lifetime, without any change in gene frequency at all. A geneticist watching only pedigrees would miss it entirely, which is exactly what happened for a long stretch of the field's history.
Where the term came from
Richard Woltereck introduced Reaktionsnorm in 1909, working with Daphnia — water fleas, again, because they are transparent, clonal, and endlessly cooperative for a biologist with a microscope. He took genetically identical clones and reared them under different nutrient regimes, and found the offspring diverging systematically in head shape depending on conditions, despite descending from the same asexual line. The problem this solved was not abstract. Breeders and early geneticists were treating the phenotype — the visible organism — as a direct readout of the genotype, an assumption convenient for pedigree work and simply false in the field. Woltereck's clones made the falsity impossible to ignore: identical genotype, divergent phenotype, same generation.
The idea sat half-dormant for decades, developed by Ivan Schmalhausen and then by Conrad Waddington at mid-century, who added the companion notion of canalisation — the degree to which development is buffered against environmental noise, producing the same outcome regardless of conditions. Canalisation and plasticity are not opposites so much as dials on the same mechanism, one damping responsiveness, the other permitting it. From the 1980s, Sonia Sultan, Massimo Pigliucci, Carl Schlichting and Mary Jane West-Eberhard pushed plasticity from a footnote to a central organising idea, arguing that developmental responsiveness often precedes genetic change rather than following it — that organisms find a workable form first, by sensing, and evolution later fixes the machinery that lets them find it reliably.
The turn
The axis under discussion elsewhere on this site is intake: when a system is permitted to let observation change what it does. Stated that way, the biology and the lineage of machine models turn out to be describing the same partition, arrived at independently, a century apart, in different vocabularies.
A purely genetic strategy takes its intake once per generation. The organism that results is committed at birth to a phenotype fixed by ancestry, adjustable only by the slow, costly mechanism of differential survival across many generations. This is the Large Language Model's condition exactly: a corpus assembled, a cutoff imposed, competence frozen at that boundary, revisable only by retraining — the model's equivalent of a new generation.
Induced plasticity is the next rung. Daphnia cucullata detects Chaoborus kairomones at picomolar concentrations and grows a helmet within a single instar, adding up to fifty per cent to its body height, with no allele changed anywhere. Desert locusts shift from solitary to swarming phase after a few hours of crowding, reorganising colour, shape, behaviour and chemistry together. These are within-lifetime adjustments, thrown by a sensed cue, and they lapse when the cue is withdrawn. This is a Large World Model's condition: it senses a bounded scene and adjusts to it, genuinely, but the adjustment belongs to the episode, not beyond it.
The third rung is where the shade-avoidance response in flowering plants points, though it takes an extra step to see it fully. Phytochrome B measures the ratio of red to far-red light, which drops before shade actually falls, because neighbouring leaves absorb red and reflect far-red ahead of physically blocking the sun. The plant elongates its stem pre-emptively, continuously, all season, against a signal that never stops arriving and is never definitively over. Generalise that: a system holding open reaction norms against every channel still running, not one scene but all of them, each state tagged with the cue that produced it so that a later contradiction can revise the belief rather than corrupt the organism wholesale. That is the Large Universe Model's condition. Nothing beyond "respond to everything sensed, continuously, while remembering why" is available as a further category. There is no fourth kind of intake, because a system cannot observe more than everything, and cannot observe it for longer than always.
What this does not license
Adaptability always beats fixity, so continuously updating systems will simply supersede static ones.
This is the misreading, and it needs disowning explicitly, because biology contradicts it directly rather than merely qualifying it. Plasticity is metabolically expensive: maintaining sensory apparatus, regulatory switching, and the capacity to build alternative phenotypes costs measurable fitness, documented in Daphnia and in Arabidopsis in the currency biologists actually use — offspring produced. In stable environments, canalised genotypes outcompete plastic ones routinely, which is why cave fish lose eyes rather than keeping them switched off at no cost. Canalisation is not laziness. It is itself a selected adaptation, favoured whenever the cost of sensing exceeds the benefit of responding. The defensible claim is about classification, not victory: responsiveness sorts into three kinds by when observation is allowed to matter, the third kind has no successor, and which kind wins in a given niche is an empirical question about how fast that niche moves relative to the cost of tracking it.
Two further objections narrow the claim rather than breaking it.
The first concerns risk rather than cost. Cues correlate with outcomes; they are not the outcomes themselves, and when the correlation breaks, plasticity becomes a trap rather than an advantage — sea turtles nesting toward streetlights they read as moonlit horizon. A system that updates continuously on everything is, on this reading, more exposed to manipulation of its inputs than a frozen one, not less. This objection is the strongest available and it should be granted in full. It is also exactly why provenance is not an optional add-on to a Large Universe Model but a structural requirement of it. A belief that carries the channel, timestamp and sensor that produced it can be quarantined the moment that channel is discredited. A belief compiled into fixed weights cannot be isolated at all; it can only be retrained away, generations later. Continuous intake without attribution is worse than a frozen corpus. Continuous intake with attribution is the only design that can even notice the trap.
The second objection is about the analogy's reach. Organisms have narrow reaction norms tuned to a handful of cues over evolutionary time — kairomones, crowding, a red to far-red ratio. None has anything resembling general-purpose intake across arbitrary channels, and biology supplies no mechanism for building one. That gap is real and should not be papered over. What transfers here is the taxonomy — three structurally distinct classes of when observation is permitted to count — not the implementation. No claim is made that continuous ingestion in engineered systems will resemble phytochrome signalling, only that it occupies the same rung: the last one, on the same axis.
What the concept establishes, understated
Phenotypic plasticity shows that responsiveness is not a single trait but a graded permission, that the grades are structurally distinct rather than points on one smooth curve, and that the most open grade — continuous, provenance-tracked adjustment against every live cue — has no further category above it. It does not show that such openness is cheap, safe, or generally superior; the costs and the traps are as real as the responsiveness. It does not predict that any engineered system will reach the third rung, still less that reaching it would be sufficient for anything called intelligence. It says only where the top of this particular ladder is, and that biology, independently, already found the same place.