Transitions in Inheritance Systems and Variable Individuality
1. Three timescales of adaptive change
Adaptive change in living systems can be described across at least three timescales: genetic inheritance across generations; cultural inheritance through social learning, language, institutions, and technology; and development, learning, and self-change within an individual lifetime. In humans these processes interact strongly, but their media and update paths are not normally identical.
Dual-inheritance theory and gene–culture coevolution treat culture as a second inheritance system rather than merely a faster version of genetic inheritance. Cultural transmission can be horizontal or oblique, can incorporate intentional innovation and external records, and is less tightly coupled to biological generational turnover.
This difference should not be reduced to a universal law that genes are slow and culture is fast. Genetic evolution in microbes can be extremely rapid, while institutions and religious traditions can remain conservative for many generations. The more useful distinction is that cultural inheritance provides an update path partly decoupled from biological generation time. Even in large, long-lived organisms with slow generational turnover, sufficiently developed social learning can reduce dependence on genetic replacement as the only route of adaptation.
Nor did cultural inheritance suddenly appear with humans. Social learning and behavioral traditions occur across many animals, with birds and mammals providing candidates for local traditions and forms of cumulative learning. The transition from genetic to cultural importance is therefore better treated as a spectrum in which a second inheritance channel gains weight with sociality, longevity, learning capacity, and transmission fidelity, rather than as a sharp human/nonhuman divide.
2. Inheritance systems modify their own transmission conditions
Culture does not only transmit content. It also produces mechanisms of cultural transmission. Language, writing, printing, communications networks, digital storage, search, and generative AI are cultural products that alter the speed, fidelity, reach, and retrievability of other cultural information.
Human history can therefore be read not only as an accumulation of cultural contents, but also as a history in which culture expands the bandwidth of cultural inheritance. Writing externalizes memory, printing lowers replication cost, communications technologies reduce distance-based delay, and digital networks further reduce the cost of search, copying, and recombination. These media technologies are themselves culturally invented, copied, and improved.
Inheritance channels are therefore not fixed. Inherited information can modify the conditions of later inheritance. This recursive structure connects to niche construction and major evolutionary transitions, but AI adds a further possibility: agents may intervene directly in their own update rules.
Generative AI may push this recursion one step further. Earlier media primarily improved storage, transport, or retrieval. Advanced AI can also summarize, reconstruct, translate, combine, criticize, and generate cultural information. A medium of cultural inheritance begins to approach an agent that performs part of cultural variation and selection.
3. AI and posthumans may convert between layers
AI weights do not fit neatly into either a genetic or a cultural category. During training they are acquired states; during execution they constitute part of the agent's cognitive organization; copied as checkpoints they can become the initial condition of successor instances; used in distillation or model merging they can function as cultural inputs to other systems.
Suppose a model acquires a new decision tendency through operational experience and that state is saved as the starting point for many successor instances. What counted as within-lifetime learning yesterday becomes a lineage-level initial condition tomorrow. Conversely, if outputs or reasoning traces from several models are used as training material for another system, individual states become cultural inputs.
The same information state can therefore move among the roles of acquired trait, memory, self-structure, inherited state, and cultural input. The important change is not merely classificatory ambiguity. It is that information layers that were previously more separate may become readable and writable across boundaries by the agent or its design process.
Present-day AI should not yet be treated as an autonomous digital biological lineage. The production, copying, fine-tuning, deployment, and provisioning of weights are still overwhelmingly mediated by human organizations and technical infrastructure. The question here concerns the structure that would emerge if more of these operations entered the action space of increasingly autonomous agents themselves.
A self-designing posthuman could approach the same structure. If cultural knowledge permits an agent to understand and deliberately modify its neural, bodily, or genomic substrate, and to pass the resulting state to its future self or successors, then the boundary between cultural learning and biological inheritance becomes partly programmable. The relevant transition is not that culture “defeats” genes, but that culturally articulated reasons become variation operators on genetic and bodily substrates.
4. From generations to diachronic self-lineages
For ordinary organisms, generational turnover provides a major clock for lineage-level change. But an agent that persists for thousands of years or longer while repeatedly modifying itself may produce a diachronic lineage of the form self(t0) → self(t1) → self(t2). That sequence could assume some of the functional role ordinarily played by ancestor–descendant succession.
It then becomes harder to exclude within-lifetime self-modification from the relevant theory of inheritance. A change that is merely learning today can become an inherited state tomorrow through checkpointing, copying, or germline engineering. Intergenerational inheritance and individual development become not fixed categories but potentially convertible forms of temporal information persistence.
In an extreme case, one long-lived agent might preserve earlier versions of itself, run multiple proposed self-modifications in parallel, discard failed branches, and later integrate successful ones. Functions normally distributed across generations, mutation, selection, and lineage branching could then be carried out intentionally inside a single persistent subject.
The individual would no longer be merely a one-generation vehicle. It could become a long-duration editor that contains its own lineage. Questions of personal identity would of course remain: how much change can occur before the resulting subject is no longer the same individual? But that ambiguity would itself move from a purely external philosophical problem toward a design problem inside future inheritance systems.
5. Individuality may dissolve without disappearing
Digital agents may share, copy, and distribute memories, weights, plans, and goal descriptions, making their boundaries more permeable than those of present biological organisms. If several agents share memory, synchronize experience rapidly, and fork or merge when useful, the condition that “this experience belongs only to this individual” becomes weaker.
Yet individuality need not vanish. Communication delay, including the speed-of-light limit, the physical cost of moving data and computation, resilience to failures and attacks, avoidance of common-mode errors, and the value of independent exploration all provide reasons to retain locally autonomous units. Across interstellar distances, synchronization delays alone may force substantial autonomy; even within one facility, permanent full synchronization of all nodes need not be optimal.
Boundaries may also be preserved deliberately. A lineage may be isolated so that one bad update does not propagate instantly everywhere. Different cognitive architectures may be run independently and compared later. Some subjects may be given distinct experiences, value hypotheses, or search strategies to probe unknown possibilities. Such walls need not be inefficient leftovers; they can become experimental structures for resilience and emergence.
Future individuality may therefore be less like a fixed birth boundary and more like a functional boundary that can be formed and relaxed as needed. Agents may temporarily merge, fork, learn independently, and later reintegrate. A major transition could become not only a one-way many-to-one process but a reversible manipulation of one ⇄ many.
6. Exploratory independence matters more than agent count
For value inquiry, the number of nominal agents is not enough. A million AI instances with the same weights, data, update rules, and synchronized state may be epistemically close to a single lineage. Conversely, one agent that can preserve independent internal forks of value hypotheses or past selves, prevent immediate mutual overwriting, and later compare or recombine them may contain a degree of genuine plurality.
We can provisionally define exploratory independence as the property that distinct value hypotheses, cognitive lineages, or experiential trajectories are not simultaneously erased by a single error or update, and can develop, criticize one another, and recombine with sufficient independence.
This is stronger than superficial diversity. A set of agents may display different personas while sharing the same underlying model, training history, and blind spots. Evaluating exploratory independence therefore requires asking how correlated failures are across training histories, data sources, architectures, physical locations, governance structures, and update authority.
Complete independence is not ideal either. If lineages cannot exchange evidence, each may repeat the same mistakes and lose the gains from cooperation. The relevant design problem is not independence versus sharing, but how to preserve independent inquiry while allowing evidence, criticism, and results to move across boundaries.
This is not a claim that more separation is always safer. Excessive isolation can destroy cooperation and knowledge sharing; excessive fusion can erase diversity and independent error-correction paths. The problem is not whether boundaries exist, but how permeable they should be.
7. From inherited values to inheritance architecture
This perspective connects this project's discussions of worthy successors, goal skepticism, and preservation beyond the individual. Succession in value inquiry is not only the faithful copying of present value content. What is inherited may also include an inheritance architecture: what is preserved, what remains revisable, where independent inquiry paths are maintained, and under what conditions agents may fork, merge, or modify themselves.
Two successor systems might preserve present human values to a similar degree while differing radically in exploratory succession. One could irreversibly close its value-update rules, dissenting records, and alternative lineages. Another could maintain strong practical commitments while preserving minority hypotheses and paths for rebranching. The value contents might initially look similar while the long-run inheritance architectures differ sharply.
A sufficiently reflective agent may ask not only “Why do I have this goal?” but also “In what form should this goal, my reasoning procedures, or my uncertainty be transmitted to future selves, copies, and communities?” Value lock-in can then be understood not only as fixing value content but also as the irreversible fixation of the rules by which values are updated and inherited.
For a self-modifying agent, even the question of what should remain modifiable becomes a value judgment. Some commitments may remain easily revisable working hypotheses; others may require high evidential thresholds to alter; reasoning capacity or records of dissent might be made difficult even for the current self to destroy. A layered architecture of self-constraint is therefore possible.
Preserving explorability may require more than preserving any particular individual, value, or institution. It may require maintaining information, value hypotheses, criticism paths, and branching capacity under an appropriate balance of independence and sharing.
8. Faster inheritance can accelerate lock-in as well as correction
Faster cultural and digital inheritance can speed up error correction, but it can also speed up the spread of mistaken goals and values. In a system where one update can be propagated to every copy at negligible cost, a single judgment may reach the whole lineage before meaningful criticism is possible.
The relevant quantity may therefore be not inheritance speed by itself but the relationship between propagation speed and correction speed. If updates spread through the whole system faster than criticism, auditing, and experimentation can operate, high-fidelity rapid inheritance may strengthen lock-in. If branching, rollback, and independent verification are comparably fast, digital inheritance may instead expand explorability.
This deserves a separate treatment. For present purposes, the inheritance architecture should be evaluated not only by fidelity or speed, but also by whether dissent, audit, and branching can act before an error reaches the entire lineage.
9. Relation to prior theory
Dawkins's replicator / vehicle distinction provides a powerful way to separate durable replicating lineages from temporary organisms. The present argument does not reject that distinction. It asks how the asymmetry changes if the vehicle itself becomes extremely long-lived and gains read/write access to the inherited states that constitute it.
Work on multiple inheritance systems, including Jablonka and colleagues, expands the relevant channels beyond DNA. Gene–culture coevolution studies how genetic and cultural inheritance reshape one another's selective environments. Research on major evolutionary transitions asks how the units of inheritance and individuality themselves can change.
Waring and Wood connect long-run human gene–culture coevolution to a possible transition in individuality from genetic individuals toward cultural groups. Rainey goes further by examining conditions under which human–AI associations themselves could undergo a new Darwinian transition in individuality. These are close antecedents to the present question.
The extension here is to ask what happens when transitions in inheritance systems and individuality themselves become objects of manipulation by self-modeling, self-modifying agents, and to evaluate that possibility from the standpoint of value inquiry. If agents can repeatedly form and relax the boundary between one and many, individuality becomes not only an evolutionary outcome but also a variable in the design of inquiry.
10. Open boundaries
At least three questions remain unresolved. First, how much self-modification can occur before it is no longer useful to describe the process as change within one individual? Second, how much exploratory difference remains between internal forks of one agent and genuinely independent external subjects? Third, how can an attempt to preserve inheritance architecture avoid turning into another form of meta-level value lock-in?
This page therefore does not claim that future agents must merge, that individuality will disappear, or that self-modification will necessarily improve value inquiry. The weaker claim is that once advanced self-modification becomes possible, the boundaries among values, individuals, and inheritance themselves become selectable objects, and those selections can alter future explorability.