Symbiosis and horizontal gene transfer: why continuous ingestion follows

The concept, on its own ground

Inheritance, in the classical picture, runs one way. Genes pass from parent to offspring, mutation supplies variation, selection sorts it, and a lineage's future is bounded by what its own genome already contains. Call this vertical descent. It is the tree of life as Darwin drew it: branching, but never touching.

Two phenomena break that picture. Symbiosis is durable partnership between unlike organisms, close enough and long enough that each depends on the other for functions it cannot perform alone. Horizontal gene transfer is the movement of genetic material between organisms outside the parent-to-offspring line — conjugation, transformation, transduction, viral insertion, outright engulfment. Neither is exotic. Bacteria conjugate plasmids across species boundaries in hours. Free-living cells were captured whole and retained as organelles. In every case, the mechanism that matters is the same: capability enters a lineage from outside it, fully formed, not built up by mutation and selection acting on that lineage's own material.

This matters because it changes what the past can tell you about the future. Under vertical descent, in principle, you can look at a genome and bound what it might become; the raw material is there, however slow the sorting. Under horizontal transfer, you cannot. The glassy-winged sharpshooter's competence at building the eight amino acids it cannot synthesise is not encoded in the insect's own genome at all — it is held by Sulcia muelleri, an endosymbiont bacterium whose genome has shrunk to roughly 245 kilobases, stripped of nearly everything except those pathways. Sequence the insect alone, however deeply, and you will not find it. Inheritance, once you admit transfer, is a network, not a tree.

Origin: two problems, one answer

Lynn Margulis proposed in 1967 that mitochondria and chloroplasts were once free-living bacteria, engulfed and retained rather than evolved in place. The idea was resisted for over a decade — it implied that eukaryotic cells were composite organisms, their core machinery of a different ancestry from their nucleus. Sequencing eventually vindicated her outright: the mitochondrial genome is recognisably that of an alpha-proteobacterium, reduced but unmistakable.

Independently, in the late 1950s, Japanese researchers investigating dysentery outbreaks found something stranger. Shigella strains were acquiring resistance to multiple antibiotics simultaneously, and the resistance was passing to unrelated E. coli in the same patients — not through breeding, which bacteria of different genera do not do, but through direct transfer of what came to be called the R factor: a conjugative plasmid, a self-contained package of DNA that moves cell to cell across species lines.

Both findings solved the same kind of problem: a trait that existed, observably, but that no reconstruction of the organism's own ancestry could explain. The response, in both cases, was to abandon the tree as a complete model and admit a network alongside it. For prokaryotes especially, that network is now understood to be the dominant mode of genetic change, not a curiosity at its edges.

The turn

Set biology aside for a moment and ask a narrower question: where does a computational system's capability come from, and what does that origin imply about its limits?

A Large Language Model's capability is bounded by its training corpus. Whatever it can produce was present, in some distributed and recombinable form, in the text fixed at a cutoff date. Better prompting, longer context, more compute at inference — these are ways of reading the same inheritance more cleverly. They are, in the biological sense, vertical: improvement within a lineage whose material stopped growing at a fixed point. This is not a criticism of the corpus. It is a statement about what kind of process is running.

A Large World Model adds something the corpus never had: a scene, present now. Sensing supplies the actual gait of this patient, the actual load on this bridge, in the moment — material that no amount of re-reading the training data could produce, because it did not exist when the data was fixed. That is acquisition from outside the lineage. But it is bounded by the episode. When the scene ends, the channel closes, and the system falls back to what it was given plus what it just saw.

The step to a Large Universe Model is to notice that the mechanism — acquisition from outside the closed inheritance — need not be tied to a single episode at all. Keep every stream open. Let any arriving observation revise a belief formed elsewhere, at any time, not only during a bounded scene. Attach provenance to every belief, so the origin of a graft is traceable and, if the source turns bad, reversible. That is not a bigger world model. It is the generalisation of the acquisition mechanism itself, freed from the episode that housed it.

This is where horizontal gene transfer becomes more than an analogy. It is the demonstration, at evolutionary scale and over billions of years, that the mechanism works — that a lineage can survive, and occasionally be transformed by, capability that arrives unpredictably from outside it, provided there is machinery to accept, integrate, and purge such arrivals. Continuous intake with revisable, provenanced belief is that machinery, restated for a system that reasons rather than replicates.

What must not be concluded

The obvious misreading is that if capability can arrive from anywhere, the correct policy is to ingest indiscriminately — more input, always, is better. Biology refutes this on its own terms. Bacteria maintain restriction-modification systems and CRISPR arrays for the specific purpose of rejecting foreign DNA. Unrestricted transfer is not symbiosis; it is infection. The lesson of horizontal gene transfer is not that gates should be open. It is that where they are opened, there must be machinery to trace what came through and revoke it if it turns out to be harmful. A Large Universe Model's claim is about revisability under arrival — the capacity to accept, trace, and retract — not about volume of intake.

Objections, taken seriously

Most horizontally transferred genes are neutral or harmful and get purged. The successful cases — mitochondria, resistance plasmids — are survivorship bias across immense stretches of time. A system that ingests everything continuously mostly ingests noise.

This is correct, and unflattering to any romantic reading of the concept. Estimates suggest the large majority of horizontally acquired sequence in bacterial genomes is lost again within a few million years. But the conclusion drawn from that fact should not be to close the intake — that sets the hit rate to exactly zero. It should be to make purging cheap. Provenance is the cheap-purging mechanism: a belief tagged with its source can be retracted the moment the source is discredited, without unwinding everything built downstream of it. Biology pays for purging with death, generation after generation. A system with provenance pays with bookkeeping.

Vertical descent produced almost all observable complexity. Horizontal transfer is a minority contributor outside prokaryotes. If most capability comes from within a lineage, scaling the corpus is the higher-return strategy, and continuous intake is a side channel.

This narrows the claim, and honestly so: most improvement, most of the time, will indeed come from working harder on what a system already has. But the frequency of exogenous events is not the right measure of their importance. The rare events include the origin of the eukaryotic cell, oxygenic photosynthesis, and the placenta — the last built partly on syncytin-1, a fusion protein that is, in origin, the envelope gene of a retrovirus that inserted itself into the mammalian genome some 25 million years ago. Placental mammals depend on a capability captured from a pathogen. Rare and decisive are not opposites here. Continuous intake is insurance against tail events, not a claim about the median case — and insurance that is never needed still needs to have been bought.

Calling continuous intake terminal is unfalsifiable. Any future sensor or acquisition method can just be relabelled as more of the same category, protecting the claim from ever being wrong.

This one has teeth, and the honest answer is to specify what would refute it. A fourth mode would have to be neither a fixed prior record, nor present sensing, nor an open, revisable stream — some acquisition method with no temporal relation to the observer at all. No candidate has been proposed, because those three exhaust the temporal structure of observation: before, now, and continuously as it happens. A new sensor is a new stream joining the open architecture, not a new relation to time. The real open question, and it is genuinely open, is whether continuous intake is governable — whether trust and provenance can scale to millions of arriving claims without collapsing into either paralysis or credulity. That is where the position can fail.

What the concept establishes, and what it does not

Horizontal gene transfer establishes that a lineage's future is not contained in its past whenever transfer is possible, and that surviving this fact requires machinery for acceptance and purging, not indiscriminate openness. Carried across to computational systems, it establishes that no amount of internal extrapolation over a fixed corpus anticipates capability arriving from outside it, and that an architecture built to survive such arrival — open streams, revisable belief, provenance — has no further rung above it on the axis of intake. It does not establish that such an architecture is easy to govern, that most gains will come from it rather than from better use of what is already held, or that the biological analogy holds beyond this one structural point. The claim is narrow by design: not that ingestion should be total, but that revisability under arrival is the last move available on this particular axis.