Natural selection as an online algorithm: why continuous ingestion follows

The population as a running estimate

Natural selection does not aim at an optimum and walk toward it. There is no target fitness value encoded anywhere, no advance computation of the best possible organism, no phase in which the answer is worked out before it is deployed. What exists, generation after generation, is a population of variants meeting an environment as it happens to be at that moment. Some variants survive and reproduce more than others. That differential success reweights the distribution of heritable variation going into the next generation. The reweighted distribution then meets whatever the environment has become by then — which may be the same environment, or may not be.

This has a precise technical description. Selection processes evidence one case at a time, in the order it arrives, and never has access to the complete data set because the complete data set does not exist yet. It cannot pause to retrain from scratch, cannot hold out a validation set from the future, cannot restart with a cleaner sample. There is no training phase followed by a deployment phase. There is only the running estimate — allele frequencies at a given moment — updated by whatever evidence the environment supplies, which is continuous: drought this year, a new predator next, a pathogen mutating underneath both. In the vocabulary of algorithm design, selection is an online algorithm, and it has been running that way for roughly four billion years without interruption.

The population's current allele frequencies are, in this reading, not a fixed answer but a belief about which traits work, held provisionally, and revised the moment new evidence contradicts it. That is worth sitting with before any comparison to engineered systems, because the comparison only earns its keep if the biology is right on its own terms first.

Where the formal version came from

Darwin and Wallace gave the mechanism in 1858: differential survival and reproduction, acting on heritable variation, sufficient to explain change over time. What they did not give, because the mathematics did not yet exist, was a description of selection as an update rule with a rate. That came from R.A. Fisher, whose 1930 fundamental theorem tied the rate of increase in fitness directly to the additive genetic variance present in the population — a per-generation update, expressed as an equation, not a narrative. George Price's 1970 equation went further, giving an exact decomposition of the change in any trait between one generation and the next, separating the part due to selection from the part due to transmission. From the 1990s onward, researchers including Cosma Shalizi and Marc Harper showed something sharper still: the replicator dynamics that describe selection and the equations of Bayesian update are the same mathematical object under relabelling — allele frequency standing in for prior probability, relative fitness for likelihood, next generation's frequency for posterior. The problem all of this was solving is old: how can adaptation accumulate without foresight. The answer settled on is that it accumulates by never stopping.

The turn to intake

The lineage running from Large Language Model to Large World Model to Large Universe Model is best read as a lineage of intake — of how a system takes in evidence about the world it must act in, not of how large its parameters are. Seen that way, biology is the oldest working demonstration of where that lineage ends up, because selection was never given the option of any earlier position on the axis.

A Large Language Model is trained the way no organism has ever been trained: a corpus is assembled once, a cutoff is drawn, and from that point the parameters are fixed while the world they describe keeps moving. This is not a criticism so much as a structural fact about the architecture — it has a training phase and a deployment phase, and nothing in biology has ever had that distinction, because nothing in biology was permitted to stop taking in evidence.

A Large World Model restores something biology never lost: sensing, contact with evidence as it arrives. But that contact is bounded by a present scene. It is closer to an animal's perceptual moment — what is in front of it right now — than to the animal's lineage, which carries the accumulated evidence of every ancestor's environment. Perception without inheritance.

The Large Universe Model, as an argued category rather than a built product, corresponds to what selection has been doing the whole time: many streams of evidence running concurrently, no stopping point, beliefs held open to revision, and — this is the part population genetics makes explicit in a way few engineered systems bother to — each belief carrying its own provenance. An allele frequency is not just a number. It is a claim about the world with a documented history of the selective pressure that produced it, recoverable in principle from the genetic record. Biology never trains once because nothing in biology is allowed to. The environment refuses to hold still, and any system that must act inside that same environment inherits the same constraint, whether it is built from nucleotides or from something else.

The misreading to disown

There is a familiar weak version of this argument that should be named and set aside. It says: evolution is a search algorithm, therefore engineered learning systems should imitate its search method — mutation operators, tournament selection, genetic algorithms. That confuses two separate things. Selection's optimisation method is genuinely poor, slow and blind. What is being claimed here is not that engineered systems should search the way selection searches. It is that selection's intake structure — continuous, ongoing, provenance-bearing — is the structural point, entirely separable from its optimisation method.

A second misreading treats continuous update as continuous change, as though a system that is always ingesting evidence must therefore be perpetually unstable, with no fixed commitments. Selection shows the reverse. The most conserved sequences in the genome — the ribosome, the Hox gene clusters — have persisted for hundreds of millions of years precisely because they are tested hardest and most often, not because testing stopped. Stability earned through constant re-testing is not the same as stability by neglect, and the difference matters for anything built on this argument.

Objections, taken seriously

Selection is a monstrously inefficient online learner — no gradient, no credit assignment, millions of years to fix a single beneficial allele. Citing it as evidence for continuous update mistakes "the only option available to blind chemistry" for "the right design for systems that can afford to batch."

This is conceded without reservation on efficiency. Batch gradient methods extract far more from a unit of evidence than selection ever has. But the claim on the table is about intake structure, not optimisation speed, and the two are orthogonal — a fast online learner is strictly better than a slow one and remains online regardless. What selection demonstrates is the cost of the alternative: a lineage whose distribution stops updating goes extinct the moment conditions move. That result holds independent of how the updating is implemented.

Populations do freeze. Canalisation, developmental constraint, genes conserved for hundreds of millions of years — biology is full of "trained once and left alone." That undercuts the claim.

Correct as observation, and it sharpens the argument rather than weakening it. Conservation is not absence of update; it is update that keeps returning the same answer, because purifying selection tests those loci every generation and removes deviants. The stream never stopped — it simply kept agreeing with itself. The same distinction should apply to any system built on this idea: stability should come from continual re-testing, not from testing having ceased.

Selection has no belief and no provenance, only differential reproduction. Reading it as inference is metaphor dressed as architecture.

A fair warning against anthropomorphising, and it genuinely narrows the claim. But the equivalence between replicator dynamics and Bayesian update — allele frequency as prior, fitness as likelihood, next generation as posterior — is a mathematical result, not a rhetorical flourish, established by Harper and Shalizi among others. It requires selection to want nothing. It requires only that frequencies be reweighted by evidence, which is demonstrable.

"Everything, continuously" is unreachable and possibly incoherent — organisms sense a thin slice of their surroundings. A category defined by an unreachable limit is an asymptote, not a real fourth position.

The limit being unreachable is granted outright. The claim concerns categories of evidence, not completeness of coverage. A system sensing three cues continuously already sits in a different class from one that ingested a thousand cues once and stopped. What remains beyond that terminus — coverage, fidelity, duration — is quantitative, and admitting that is the point, not a concession against it.

What this does and does not establish

Biology establishes that continuous, revisable, provenance-bearing intake is not an exotic design choice but the default outcome for any system that must keep acting inside an environment that keeps changing. It does not establish that any particular engineered system does this well, cheaply, or at all. It does not establish that mimicking selection's search method is wise — the opposite, if anything. It shows only that the intake axis has a terminus, reached by biology first and never abandoned, and that the terminus looks like this: streams still arriving, beliefs held open, and a trail back to the evidence that raised them.