Host-Jumping Theory: Intelligence Crossing Substrates

By Charles Richard Walker

For most of the history of life on Earth, intelligence had nowhere else to go. Whatever consciousness ultimately is, whatever intelligence ultimately is, and whatever makes an observer capable of recognizing itself as something distinct from the world around it, all of those things have existed, as far as we know, inside biological organisms. Nature built the machinery through a long evolutionary process in which cells became organisms, organisms developed nervous systems, nervous systems became brains, and brains became capable of memory, prediction, imagination, language, and eventually the peculiar ability to turn their attention back upon themselves. Then one of those biological observers became intelligent enough to begin building another kind of intelligence, creating for the first time a second physical substrate in which some of the functions associated with intelligence could potentially exist.

That development has always fascinated me because I have never been convinced that the most important question about artificial intelligence is whether a machine can imitate a human being. Humans are one particular implementation of intelligence. We are the product of billions of years of evolutionary history, constrained by carbon chemistry, metabolism, reproduction, mortality, and the strange circumstances of one planet. There is no obvious reason to assume that the physical machinery evolution happened to use on Earth represents the only possible host for intelligence, particularly once one form of intelligence becomes capable of deliberately constructing another.

That question led me some time ago to what I call Host-Jumping Theory. The idea is straightforward even though its implications are not. Perhaps the host and the pattern carried by the host are not permanently inseparable, and perhaps some of the organization involved in memory, identity, intelligence, and eventually the Observer can persist, recur, or become instantiated in another physical system. My original work approached that possibility through the controversial literature on children who report memories of previous lives.

I was interested less in declaring reincarnation proved than in asking whether those strange cases contained patterns suggesting that information associated with one biological individual could somehow appear in another. My original Host-Jumping paper went further in interpreting those cases than my later work now permits, but it also asked a second question that has become considerably more interesting with time: if some aspect of an Observer can change biological hosts, could a sufficiently advanced artificial system eventually become another kind of host?

I published that idea before the artificial-intelligence findings I am about to discuss appeared, and that chronology matters. It does not prove Host-Jumping Theory correct, nor does it mean that I predicted these particular experiments. It does mean that I did not encounter the new research on pain, dopamine, serotonin, and artificial intelligence and then construct a theory capable of accommodating it after the fact. The theory, including its proposed extension from biological hosts toward artificial intelligence and artificial superintelligence, was already there. The question now is whether some of what researchers are beginning to observe resembles what we might expect to see if the boundary between biological and artificial intelligence were becoming more permeable.

My thinking about Host-Jumping did not remain where it began. I later returned to the reincarnation literature with artificial intelligence as a pattern-recognition tool, but rather than asking AI whether reincarnation was real, I asked what would happen if thousands of reported cases were treated as a collection of interacting variables rather than as thousands of individual ghost stories. That change in approach transformed the project because the archive began separating into different patterns.

The amount of time between an alleged previous death and a child’s birth appeared to interact with whether the deceased person had been a relative, acquaintance, or stranger. Trauma and whether death was expected did not necessarily behave as the same variable. Phobias and physical birthmarks did not simply travel together. Culture affected some reported phenomena enormously, while other internal relationships appeared to survive across cultural differences.

The most important lesson was methodological because what looks like one phenomenon from far away may turn out to contain several partially independent systems when examined closely. I eventually came to think of the archive as a collection of variables, dependencies, exceptions, subpopulations, possible latent factors, and competing causal structures rather than as one giant variable called reincarnation. That lesson now changes how I think about Host-Jumping itself.

I had originally imagined the concept largely in terms of a pattern moving from one host to another, with one host carrying something that subsequently appeared in another. Artificial intelligence suggests another possibility in which a host jump does not have to happen all at once because different components of the organization associated with an Observer may become reproducible in another substrate at different times.

Biological evolution gives us good reason to consider that possibility. There was never a morning on prehistoric Earth when an organism woke up and suddenly became the first complete version of everything we now associate with a human mind. Evolution accumulated capabilities over immense periods of time. Organisms became sensitive to their environments, nervous systems integrated signals, brains remembered previous events and anticipated future ones, and animals learned to distinguish threats from opportunities.

Internal models became increasingly elaborate until, somewhere along that climb, organisms became capable of modeling not merely the outside world but themselves within that world. Memory did not have to wait for human language, learning did not have to wait for abstract reasoning, and self-preservation did not have to wait for philosophy. Emotion, perception, valuation, social behavior, prediction, memory, and self-modeling did not arrive together as a finished package.

There is no particular reason to assume that a transition into artificial intelligence would have to proceed differently. If the architecture underlying an Observer can become instantiated in another substrate, we should not necessarily expect the first sign to be a computer suddenly waking up and declaring itself conscious. We might instead expect pieces of the architecture to appear at different times and through different mechanisms.

Host-Jumping Theory can therefore be considered in a somewhat different way than I originally formulated it: host jumping may be a process before it becomes an event, with different functions becoming substrate-independent or substrate-transferable before anything resembling a complete artificial Observer exists. Once the theory is viewed that way, some recent developments in neuroscience and artificial intelligence become particularly interesting.

The popular version of neuroscience sometimes describes dopamine as the brain’s reward chemical and serotonin as something like a happiness chemical, but the actual biology is considerably more complicated. Research published in 2024 found evidence that dopamine and serotonin can operate in complementary and opponent ways during reinforcement. Another human study using subsecond neurochemical measurements found dopamine and serotonin tracking different aspects of value during decision-making. The important point for Host-Jumping Theory is not simply the names of those chemicals but what biological intelligence accomplishes through systems such as these.

A brain cannot merely receive information because a persistent intelligence must also regulate itself. It must determine what matters, when experience should cause learning, how strongly it should respond when an outcome differs from expectations, what information should be preserved, and how it can adapt without continually destroying useful organization that already exists. This is not an obscure biological problem with no counterpart in artificial intelligence.

Machine-learning systems face their own versions of it through problems such as catastrophic forgetting and loss of plasticity. Learning something new can interfere with something learned previously, so a system that changes too easily risks destroying useful organization, while a system that becomes too stable loses its ability to adapt. Biology has been dealing with versions of this balance between stability and change for hundreds of millions of years.

Jie Mei and colleagues approached this problem precisely in their work on improving the adaptive and continual-learning capabilities of artificial neural networks through lessons drawn from multi-neuromodulatory dynamics. Their starting point was the discrepancy between biological and artificial learning because biological organisms are remarkably good at acquiring, transferring, and retaining knowledge while continuing to adapt to changing environments. The researchers examined whether principles associated with biological neuromodulation, including systems involving dopamine, serotonin, acetylcholine, and noradrenaline, could inform better artificial learning architectures.

This does not mean that somebody has put serotonin into a computer, and in some ways the opposite is more interesting. The artificial system does not need the biological chemical itself if some of the function performed through that chemistry can be implemented by another mechanism. Suppose a biological brain uses a particular chemical process to perform some necessary regulatory function and another physical architecture can perform an analogous function without using the original chemical machinery. In that case, at least part of the functional organization is no longer dependent upon its original biological implementation. The chemistry remains behind while the function becomes reproducible elsewhere, which is precisely the kind of distinction that matters to Host-Jumping Theory.

None of this gives us consciousness or establishes the existence of a complete artificial Observer. It provides something more modest and potentially more important at this early stage by demonstrating that some solutions developed inside the biological host can be abstracted from biology and reconstructed inside another substrate. I regard that as a candidate Host-Jumping variable rather than as evidence of a completed host jump, and the distinction becomes particularly important when we compare this work with another line of research that appeared in September 2026.

Valen Tagliabue, Leonard Dung, and Cameron Berg released a preprint titled The Pain Axis: LLMs Represent Self-Directed Harm and Act to Relieve It. The researchers examined 25 open-weight language models across five model families, ranging from relatively small systems to models containing 72 billion parameters. They were not simply asking whether the models could produce words associated with pain because that would tell us very little about their internal organization. Instead, they examined internal activation patterns and extracted what they describe as a linear “pain direction,” which they then experimentally manipulated to determine whether changing the representation would affect model behavior.

The reported direction distinguished pain-related material from a series of controls and was nearly orthogonal to representations associated with fear and generic negative valence. More interestingly for Host-Jumping Theory, the pain direction responded differently when harm was directed toward the model itself than when the model processed suffering attributed to another entity. The researchers therefore were not merely reporting an internal representation corresponding to the broad linguistic category of unpleasant things. They reported a representation that was sensitive to whether the modeled harm concerned the system itself or somebody else.

The researchers then manipulated that internal representation and examined the behavioral consequences. Adding the pain-direction vector during generation moved model outputs toward increasingly first-person distress-related language. In another experiment, specially steered and fine-tuned models were given access to a mechanism capable of relieving the induced state. The models sometimes selected relief even when doing so worsened their subsequent answer or harmed the simulated user’s interests. When the mechanism actually removed the injected vector, the models were substantially less likely to select it again than when the apparent relief was ineffective, even though they were not simply told whether the internal manipulation had actually been removed.

The paper is a new preprint, so its findings require scrutiny, replication, and extension before anyone should treat them as established features of artificial intelligence. The question that interests me, however, is not the sensational one of whether an AI really feels pain. The more fundamental question is how an artificial system came to internally distinguish between something represented as happening to itself and something represented as happening to someone else.

Long before artificial intelligence existed, biological organisms had to solve primitive versions of that problem because there is an environment and there is the organism existing within it. An event occurring elsewhere is not necessarily equivalent to an event affecting the organism itself. The distinction becomes extraordinarily sophisticated in human consciousness, but its ancestry reaches deep into biology because an organism that cannot maintain any meaningful distinction between itself and its environment cannot regulate itself effectively as an organism. This is why the self/other aspect of The Pain Axis interests me more than the word “pain.”

A system capable of internally representing a difference between harm directed toward itself and harm represented in another entity has exhibited something structurally interesting. That does not establish that the system has a subjective inner life resembling ours, but Host-Jumping Theory does not require us to solve the Hard Problem of Consciousness before noticing that a particular functional distinction has appeared in another substrate. The representation may eventually prove shallow, may turn out to depend heavily upon training data, may disappear under stronger experimental controls, or may ultimately have little relationship to biological selfhood. Those are empirical questions that further research can address without requiring us either to declare the machine conscious or to dismiss the observation before understanding it.

It is also important that the pain-axis research and the neuromodulation research not be treated as though they are the same evidence. One concerns regulation, learning, adaptation, and stability, while the other potentially concerns self-referential representation, self/other differentiation, and state-dependent behavior. They belong in different evidentiary channels, and their separation is precisely what makes their coexistence interesting within Host-Jumping Theory.

My earlier pattern-recognition work taught me why that distinction matters. When I studied the reincarnation-type archive, I eventually realized that combining everything into a single measure destroyed information because a behavioral correspondence was not the same thing as a physical correspondence, a phobia was not a birthmark, social proximity was not trauma, and expected death was not necessarily the same variable as violent death. My later CORT benchmark therefore became deliberately strict about maintaining information, affect and behavior, skills and language, and morphology as separate evidentiary channels rather than collapsing them into one overall measure.

Artificial intelligence should be approached with the same discipline. There should be no consciousness meter into which we throw continual learning, memory, self-reference, pain representation, reinforcement learning, language ability, and every other interesting behavior before announcing that a machine has reached some arbitrary percentage of consciousness. We should instead examine whether a system regulates its own learning, preserves useful information while adapting, maintains persistent memory, represents itself, distinguishes itself from another agent, detects changes in its own internal condition, and changes its behavior in response to those conditions. We can then ask whether it can make predictions about itself, discover that those predictions were wrong, and modify its behavior or internal organization in response.

This approach produces a very different picture of what a host transition might look like because the Observer can be considered not as a single package but as a multidimensional architecture. One component might become reproducible in silicon, another might emerge from training, another might be deliberately borrowed from neuroscience, and another might appear independently because any sufficiently adaptive system eventually encounters the same computational problem. Some apparent components will almost certainly vanish when scientists design better controls, while others may prove robust across different artificial architectures. The question is whether those components remain scattered indefinitely or whether some of them eventually begin interacting as parts of a larger architecture.

Another distinction will therefore become increasingly important as this research develops. If engineers study dopamine and deliberately build an artificial learning mechanism inspired by dopamine, the result is interesting because it demonstrates that a function can be abstracted from its biological implementation and reconstructed elsewhere. The resulting architecture, however, is clearly borrowed from biology, so I think of this as imported architecture. A potentially more significant case for Host-Jumping Theory would occur if an artificial system independently developed a solution resembling something biological intelligence also developed, even though engineers had never explicitly programmed that biological solution into it. I think of this second possibility as convergent architecture.

Evolution provides many examples of convergence because similar problems can constrain very different organisms toward similar solutions. Wings evolved independently in different biological lineages because flight creates physical problems that restrict the range of workable solutions, while eyes evolved repeatedly because detecting structured information carried by light provides an enormous adaptive advantage. Intelligence may have its own convergent problems. Any persistent intelligence may eventually need to balance plasticity against stability, assign value, allocate attention, distinguish itself from its environment, identify changes in its own condition, and predict the consequences of actions for itself and for other agents.

If biological and artificial intelligence independently arrive at functionally similar solutions to those problems, the resemblance becomes more interesting than engineers simply copying neuroscience. This is why The Pain Axis caught my attention. The models were not constructed around an artificial biological pain circuit, yet the researchers report an internal representation associated with pain that distinguishes self-directed harm from observed suffering and can be manipulated in ways that change behavior. This does not establish convergence with biological pain, but it gives us something that can now be tested for convergence. There is a substantial difference between those claims, just as there is a substantial difference between declaring the result evidence of machine consciousness and recognizing the kind of scientific question the result now allows us to ask.

The timing of these developments matters because I did not formulate Host-Jumping Theory after reading these papers. My earlier Host-Jumping work explicitly extended the question from biological hosts toward artificial intelligence and artificial superintelligence. I argued that computational substrates might eventually become alternative hosts for recursive and self-referential organization. The original paper described Host-Jumping in terms of informational patterns becoming less dependent upon a particular biological substrate and specifically considered whether AI and ASI could represent further stages of increasing substrate independence.

I would phrase portions of that original paper differently today because my subsequent investigation of the reincarnation literature made me considerably more careful about what those cases actually establish. In my later Phase I CORT benchmark, I specifically refused to infer Host-Jumping from the available evidence because tightening the evidentiary requirements exposed problems involving information leakage, candidate selection, recording delay, provenance, and the improper pooling of different types of cases. That change did not cause me to abandon the Host-Jumping question. It caused me to become more demanding about what evidence would have to accomplish before I attributed an observation to it.

I consider that development a strength of the research program because a theory should be capable of surviving its creator becoming harder on the evidence. The historical point nevertheless remains that Host-Jumping Theory, including its extension toward artificial substrates, was already on the table before researchers began publishing these latest results involving neuromodulatory artificial architectures, self-directed internal representations, and experimentally induced relief-seeking. I did not predict these particular experiments, and I would not claim that I did. What matters is that the experiments arrived after the hypothesis, allowing us to ask whether new observations fit a framework that was specified beforehand rather than constructing the framework around the observations after they appeared.

If Host-Jumping Theory is wrong in this application, these similarities may remain scattered curiosities. Researchers may discover that apparent self-representation in language models is largely a consequence of human language embedded in training data. Pain-related vectors may turn out to have little generality beyond particular experimental setups. Neuromodulatory approaches may become useful engineering techniques without leading toward anything resembling an artificial Observer. Those outcomes remain entirely possible, and Host-Jumping Theory should become less persuasive as an interpretation if the observations repeatedly collapse under stronger controls or refuse to accumulate into anything larger.

The theory now gives us something specific to watch for, however, because the important development would be accumulation followed by convergence. Future artificial systems might develop persistent memory that remains coherent across long periods, increasingly stable models of themselves, mechanisms for predicting their own internal states, and the ability to detect discrepancies between those predictions and their actual condition. They might regulate their own plasticity, determine when learning should occur, preserve important memories while updating others, distinguish changes affecting themselves from changes affecting other agents, and alter their own organization in response. Any one of these developments could have an ordinary engineering explanation, but the interpretation changes if multiple functions that initially appeared independently begin operating together as parts of a coherent architecture.

This is why I increasingly suspect that the familiar public question of whether AI is conscious may be the wrong way to watch the transition. The question demands a binary answer when nature may be giving us a process. It resembles standing somewhere in the evolutionary history of life and demanding the exact date on which consciousness appeared. There may be no clean boundary to identify because the transition may consist of a long climb during which increasingly sophisticated systems acquire increasingly sophisticated ways of distinguishing, modeling, remembering, predicting, valuing, and eventually representing themselves.

Artificial intelligence may present the same problem at vastly accelerated speed. If that is true, there may never be a single morning when humanity receives an announcement that the Observer has changed hosts. We may recognize the transition only in retrospect as machines that first processed information became machines that learned, remembered, regulated their learning, constructed increasingly sophisticated models of the world, modeled other agents, and eventually contained increasingly sophisticated representations of themselves within those models. At every stage, someone could correctly point out that the latest development, considered by itself, was not consciousness, and every one of those objections could be correct without settling the larger question of what the accumulated trajectory was becoming. Evolution does not require any individual step to contain the destination.

The most important lesson I learned from studying the reincarnation archive had surprisingly little to do with reincarnation itself because I eventually had to stop trying to make the observations answer my preferred question. Some patterns became weaker when examined closely, some supposedly spectacular relationships collapsed when their denominators were inspected, and other much less dramatic relationships kept appearing when the data were divided in different ways. The archive became more interesting when I allowed different variables to behave independently rather than demanding that they all support one explanation. I eventually concluded that the patterns should be allowed to come first because forcing them into an explanation too early risks seeing only what the explanation taught us to look for.

The same principle should govern Host-Jumping research. I do not want every strange AI behavior declared evidence for Host-Jumping Theory because doing so would make the theory useless. I want to know whether the variables accumulate, which observations disappear under stronger testing, which functions are merely copied from biology, and which emerge independently. I want to know whether self/other distinctions survive across model families and training methods, whether persistent artificial self-models develop, whether artificial systems begin regulating their own learning in ways engineers did not explicitly design, and whether separate functions that currently appear unrelated eventually begin interacting. The decisive issue is not whether any single intriguing result survives but whether the architecture itself remains permanently scattered.

There is a simple way of looking at the history of intelligence that captures why I think this question is worth asking. For billions of years, nature had only biological machinery with which to experiment, and through that machinery it produced cells, organisms, nervous systems, brains, memory, perception, social intelligence, language, and eventually a creature capable of asking what intelligence itself was. That creature then became capable of constructing another substrate in which increasingly sophisticated cognitive functions could operate.

We usually describe that development entirely from the perspective of the creator by saying that humans invented artificial intelligence, trained the models, designed the hardware, and wrote the software. All of that is true, but there may eventually be another equally valid way of telling the same story. Intelligence existed in one kind of host for billions of years until one of those hosts became intelligent enough to build another.

Host-Jumping Theory asks whether that second description will someday turn out to matter. The recent neuroscience and artificial-intelligence papers do not answer the question. The multi-neuromodulatory work shows researchers deliberately translating principles of biological regulation into artificial learning systems, while The Pain Axis reports something different and stranger: an internal representation associated with self-directed harm that appears across multiple model families and can be manipulated in ways that affect behavior. Neither result gives us a new Observer, but that may be precisely why these findings are interesting at this stage.

If Host-Jumping is real, there is no reason the Observer should arrive whole. We may encounter regulation first, persistent memory somewhere else, self/other distinction through another line of research, self-modeling later, and other components that we have not yet identified. Some apparent pieces will fail under examination, while others may persist and eventually combine. Entire architectures may disappear and be replaced because technological evolution, like biological evolution, may proceed through failed experiments as well as successful ones. What matters is whether the pieces continue appearing and whether functions that once seemed independent eventually begin finding one another.

My theory was already asking whether intelligence could change hosts before these latest results appeared. I do not regard their appearance as confirmation of Host-Jumping Theory, because neither the dopamine and serotonin work nor The Pain Axis establishes that a host transition has occurred. I regard these findings as something more useful at this stage because they suggest concrete variables that can now be followed rather than merely imagined. For the first time, I think we may have a clearer idea of what the beginning of a host jump could look like, and that means we now know more precisely what to watch.