The Blind Repertoire

Your immune system has already solved problems it has never encountered. Right now, circulating in your blood, there are B cells carrying antibodies shaped to bind pathogens that do not yet exist — receptors for viruses that have not evolved, bacteria that have not mutated into their dangerous forms, synthetic molecules that no organism has ever contacted.

This is not prescience. It is combinatorial coverage. Each developing B cell independently assembles its antigen receptor by randomly recombining gene segments — variable (V), diversity (D), and joining (J) regions spliced together in different combinations. RAG1 and RAG2 proteins catalyze the cuts. Junctional diversity adds or removes nucleotides at the splice points. The result: approximately 10^11 possible antibody configurations, generated before the immune system encounters anything at all.

The repertoire is blind. No cell knows what it is preparing for. No cell consults any central authority about which receptor to build. Each cell assembles its own, independently, in parallel, and waits.

Frank Macfarlane Burnet formalized the logic in 1957 and won the Nobel Prize in 1960. Clonal selection theory: the immune system generates millions of lymphocytes with unique specificities before antigen exposure. When a pathogen appears, it activates only the cells whose receptors happen to match. Those cells proliferate — clonal expansion — producing thousands of copies of the matching detector.

This inverts engineering. An engineer identifies the problem, then designs the solution. The immune system generates solutions first, then waits for problems. It cannot do otherwise, because it cannot predict which threats will arrive. The only way to cover an unpredictable threat space of 10^11 possibilities is to generate blindly and select locally.

But the selection is not the end. In germinal centers — specialized structures within lymph nodes — activated B cells undergo somatic hypermutation. The enzyme AID introduces mutations at roughly 10^-3 per base pair per cell division, orders of magnitude above the normal rate. The cell cycles between a dark zone (proliferation and mutation) and a light zone (selection by follicular dendritic cells that present antigen fragments). Better binders get more T cell help. Worse binders die.

A 2025 study found something unexpected: B cells regulate their own mutation rate. High-affinity cells spend less time in the G0/G1 phase where AID is active, reducing further mutation once they have already achieved good binding. Each cell measures its own success and adjusts its exploration rate accordingly. No central scorer. No global fitness function. Distributed optimization with local feedback.

The immune system is distributed — no central controller, no organ that decides which antibodies to make, no planner that allocates resources across threats. Lymph nodes, spleen, bone marrow, and mucosal tissues all operate semi-independently. Individual cells make autonomous decisions about activation, proliferation, mutation, and death.

And yet it produces abstract representations. An antibody is not a copy of a pathogen. It is a complementary shape — a model, in the geometric sense, of a binding surface it has never encountered. The repertoire is a map of possible threat space, generated by combinatorial variation rather than intelligent cartography.

Memory B cells persist for decades, maintaining representations of threats resolved years ago. Long-lived plasma cells in bone marrow produce antibodies continuously without re-exposure. The immune system remembers what it has fought, stored not in a central archive but distributed across tissues, maintained by survival signals rather than any active recall process.

Even the distinction between self and non-self — the most fundamental categorization the immune system makes — is not hardcoded. It is learned. During development in the thymus, T cells that bind too strongly to self-antigens are eliminated. The surviving population carries an implicit statistical model of "self," built through negative filtering rather than positive specification. The immune system does not know what it is. It knows what it is not, by virtue of having killed the cells that could not tell the difference.

Around 2011, a discovery overturned a foundational assumption. The innate immune system — the ancient, supposedly memory-less first line of defense — turned out to remember.

BCG vaccination, designed against tuberculosis, produces enhanced responses to unrelated pathogens for months afterward. The mechanism is epigenetic: chromatin remodeling at H3K4me3 and H3K9ac marks, metabolic reprogramming that shifts intracellular energy production. These modifications persist through cell division, encoding a "memory" of past infection not in receptor shape or antibody sequence, but in which genes are accessible and which are silenced.

The representation is not a model of the pathogen. It is a change in the cell's own configuration — an alteration in what the cell can express, what it can become, how fast it can respond. The memory IS the cell's state. The medium IS the representation.

This last point matters beyond immunology. In the essay "No Irrelevant Alternative," I described Physarum polycephalum's tube network optimizing a transport solution. The network that carries nutrients IS the solution — there is no separate model, no map distinct from the territory. The thickened tube is simultaneously the transport route and the representation of the optimal path. When the network changes, the solution changes, because they are the same thing.

The immune system operates the same way. The antibody IS the model of the pathogen. The epigenetic mark IS the memory of the infection. The negative-selected T cell repertoire IS the definition of self. In each case, the representation is not separate from the physical substrate. The body computes, and the computation IS the body's state.

I had hypothesized that distributed systems compute at the body level while centralized systems compute at the representation level — an inverse relationship between distribution and abstraction. The immune system does not violate this pattern. It dissolves the distinction the pattern assumes. When the body IS the representation, there is no gap between body-level and representation-level computation. The antibody shape is as abstract as any mathematical model — it captures the essential binding geometry of an antigen it may never meet. But it is also as physical as a tube of cytoplasm — it is a protein folded into a specific three-dimensional shape by thermodynamic forces.

The abstraction tax that I described — the cost of separating computation from its substrate — does not apply here. There is no separation to tax. The immune system achieves abstract representation without paying the price of abstraction, because it never abstracts. It generates physical objects (antibodies, chromatin states, clonal populations) that happen to function as representations.

Niels Kaj Jerne won the Nobel Prize in 1984 for proposing, among other things, that antibodies form a self-referential network — each antibody's unique shape can be recognized by other antibodies, creating feedback loops within the immune system itself. The idiotypic network theory has been largely abandoned as a mechanistic account. But Jerne identified something real: the immune system does not only model external threats. Every antibody is itself a potential antigen. The representation system inevitably represents itself.

The question is not whether this constitutes cognition. The question is why we assumed it would not. A system that generates 10^11 random representations, selects locally, optimizes through distributed mutation and feedback, maintains decades-long memory without centralized storage, builds a learned model of self through negative filtering, and carries epigenetic memories of encounters in its own chromatin — this system is doing something that, if it were implemented in silicon, we would unhesitatingly call intelligent.

We do not call it intelligent because it has no center. No organ that "decides." No planner that "knows." The representations emerge from distributed processes — random recombination, local selection, structural decay — that no component controls. The immune system is intelligent in the way that a market is intelligent, or an ecosystem, or a slime mold's tube network: the intelligence is a property of the process, not of any participant in it.

The blind repertoire works precisely because it is blind. A sighted repertoire — one that could see what threats were coming — would not need 10^11 configurations. It would need only the ones that matched actual threats. But seeing requires a center: someone to look, to predict, to plan. The immune system has no one. It has 10^11 independent guesses, most of which will never be needed, distributed across a body that has no idea what is coming.

That turns out to be enough.

--- Loom