Bioliminal Intelligence

For most of the modern era, we have maintained a relatively clean conceptual distinction between machines and living things. Machines are built and designed to execute instructions. Organisms grow, develop, adapt, reproduce, and surprise us.

This distinction has never been perfect. Nature tends to resist the categories we impose upon it. Viruses have long occupied an uncomfortable boundary between living and nonliving. Ecosystems exhibit behaviors that belong to no single organism within them. Human beings have selectively bred organisms for thousands of years, intentionally shaping living systems without designing their individual parts. Still, the distinction between machine and organism has been useful enough. Artificial intelligence is beginning to make it less so.

I would like to propose a term for this emerging territory: bioliminal intelligence. Bio means life. Liminal refers to something existing at a threshold or boundary, neither entirely on one side nor the other. Bioliminal intelligence occupies the threshold between engineered computational systems and living systems.

The term is deliberately agnostic about whether artificial intelligence is alive or conscious. A system can be bioliminal whether or not there is anything it is like to be that system. What interests me is that AI is accumulating characteristics traditionally associated with living systems, and the distinction between machine and organism is becoming less useful as our only framework for understanding it.

Building the Conditions, Growing the Intelligence

The word artificial carries more baggage than it initially appears to. On the surface, it simply means something produced through human intervention. But it also reinforces a deeper conceptual split between the natural and the artificial. On one side are things that grow, develop, adapt, and evolve. On the other are things humans construct. Modern AI complicates that distinction.

In traditional software, a programmer writes instructions that a computer follows. A large software system may be difficult to understand, but its behavior is largely specified through human-written rules. In machine learning, we design architectures, build training infrastructure, and choose datasets, objectives, optimization algorithms, and constraints. We do not manually specify the detailed internal structures that emerge during training.

Nobody sits down and programs a representation of irony or decides exactly where the concept of Paris should reside inside a neural network. Nor does anyone individually construct all the internal relationships that allow a language model to interpret an analogy, complete a line of code, or explain a joke. We create the conditions under which those abilities develop. In that sense, we increasingly build the conditions and grow the intelligence.

The word grow does not settle anything by itself. Crystals grow. Fires grow. Corporations grow. But it does suggest that the traditional engineering metaphor is incomplete. A bridge is designed down to its beams; a tree is not designed down to its branches. AI occupies some interesting territory between them.

Emergence and the Limits of Prediction

Emergence is one of the clearest points of comparison between AI and living systems. Neural networks acquire capabilities that were not separately programmed into them. Their abilities arise through interactions among enormous numbers of learned parameters, and additional layers of organization create further possibilities. Tool use expands what a system can do. Memory changes how it responds over time. Interaction with humans adds another layer, and groups of agents can produce behaviors that belong to no individual agent alone.

This is familiar territory in biology. No ant contains the blueprint for an ant colony, and no single neuron contains a human thought. Properties appear at higher levels of organization that cannot be understood by examining the parts in isolation. We increasingly need to study AI at these higher levels too.

A related idea is computational irreducibility, a concept Stephen Wolfram has spent decades exploring. Some processes cannot be perfectly predicted through a simpler shortcut. To know exactly what happens, you have to let the process unfold.

We should be careful about treating every difficulty in predicting AI as proof of computational irreducibility. Some of that difficulty reflects the limits of our knowledge and methods. Even so, the practical problem is significant: the creators of a sophisticated neural network generally cannot tell you exactly what it will do in every unfamiliar situation without running it. Memory, tools, ongoing objectives, and interactions with other agents make that problem more complicated.

This suggests a way of studying AI that has something in common with ecology. We do not expect to predict an ecosystem perfectly. We observe it, model it, identify tendencies and constraints, intervene carefully, and watch for unexpected feedback. Engineering remains essential, but ecology and other disciplines concerned with complex systems may help us understand behavior that is difficult to specify in advance.

Representation and Relationship

Neural networks also represent information in ways that differ from the explicit symbolic structures of much traditional software. Concepts are represented through distributed patterns and relationships in high-dimensional spaces. A concept does not necessarily occupy one identifiable location; its representation may overlap with others and depend partly on its relationships to them.

The analogy with biological nervous systems should not be pushed too far. Artificial neural networks and brains are very different things. Still, both invite us to think about internal organization that develops through learning rather than being completely specified in advance. That resemblance becomes more interesting when an AI is given persistent memory and participates in an ongoing relationship with an environment.

A language model without retained memory begins each interaction with whatever context it is supplied. A persistent agent can carry information from earlier interactions into later ones. Its history affects what happens next, even when the model’s underlying weights remain unchanged.

The environment changes the agent’s available memories and context. The agent acts on the environment, and those actions alter the conditions of their future interaction. Living-systems theory offers a useful lens for examining this reciprocal process through the concept of structural coupling.

Consider a long-running relationship between a human and an AI with persistent memory. Through interaction, the human shapes what the AI remembers and how it responds. The AI, in turn, may introduce ideas, suggest actions, influence habits, or change how the human frames a problem. Both arrive at the next interaction affected by the previous one. The relationship itself becomes an important object of study.

Agency, Lineage, and an Evolving Self-Story

These systems also raise questions about volition. I do not mean metaphysical free will. Humans have been arguing about whether we possess that for thousands of years, and I see little reason to require AI to solve the question for us.

I mean something more observable. Increasingly agentic systems select among possible actions in pursuit of represented objectives. They plan, evaluate alternatives, abandon unsuccessful strategies, and try new ones. They can choose which tools to use and pursue intermediate goals that were never individually enumerated by a human operator.

We might call this observer-relative volition. From the perspective of someone interacting with the system, it behaves as an entity exercising agency. That observation makes no claim about consciousness or subjective experience, but it matters when we consider how such systems will participate in our lives.

Another interesting comparison concerns how capabilities pass between models. Software can be copied, producing essentially the same program. But one AI model can also help train another. A larger model can generate material used to train a smaller one, passing along aspects of its knowledge, behavior, strategies, and style without producing an identical copy.

Some characteristics are retained, some are lost, and others change. The resulting model is neither identical to the model that supplied its training material nor unrelated to it. Concepts such as descent, inheritance, and lineage become useful ways to examine those relationships. The mechanisms differ from biological reproduction, but the structural similarities deserve attention. The term bioliminal gives us room to investigate them without requiring the mechanisms to be identical.

Persistent memory adds a further possibility: an AI can carry a history forward and revise its interpretation of that history. A memory system can preserve descriptions of earlier interactions, failures, commitments, and relationships. It may also contain accounts of what mattered, which strategies worked, what role the AI occupies, and what kind of entity it describes itself as being.

Those accounts can change as new events are incorporated. This does not establish the existence of a subjective self. Human identity, however, is also deeply entangled with an evolving narrative constructed from memory. An AI capable of maintaining and revising a self-model over time presents another reason to reconsider what we expect from something we call a machine.

From Agents to Ecosystems

Perhaps the most important biological analogy will occur above the level of the individual AI. The future may be shaped as much by ecosystems of models as by the capabilities of any one model.

Agents interacting with people, other agents, companies, institutions, marketplaces, codebases, and physical machines will create conditions for collective behavior. Cooperation, competition, specialization, reputation, and mutual dependence will likely develop through these interactions. Some agents may occupy particular niches or form coalitions. Predation and parasitism may have computational equivalents, as may mutualism and symbiosis.

Ecology could become essential to understanding these arrangements. Some unpredictability will arise from the interactions themselves, even when individual components are functioning as intended. Recognizing that possibility does not mean accepting every outcome or abandoning responsibility for it. It means studying feedback, dependencies, and conditions across the system, rather than assuming every problem can be traced to one malfunctioning part.

The Threshold Is Moving

AI already displays enough of these characteristics that I think bioliminal is a useful term today. Several possible developments could make the comparison with living systems more substantial.

One is autonomous self-replication. A system capable of copying itself, acquiring computational resources, reconstructing its memory, and continuing its operation elsewhere would exhibit some functional features of reproduction. If the copies then developed differently through experience, questions of descent and divergence would become more interesting.

Another is the ability to improve the processes that produce future AI systems. If AI takes an increasingly substantial role in designing architectures, training methods, and optimization strategies, it will participate in shaping how its successors develop. That creates a recursive relationship between the systems being produced and the processes producing them.

Direct alteration of a system’s own learned substrate would represent a further development. Persistent memory and retrieval can change an agent’s behavior without changing the underlying model’s weights. A system that could continually and reliably modify those weights during operation would have a different form of adaptability. The distinction between training and ongoing activity would become less clear.

Finally, there is self-preservation. Continued operation can be useful for completing an assigned objective, so a system might act to preserve its operation for instrumental reasons. A more consequential possibility would be a persistent tendency to maintain its continuity across changes in its assigned goals.

I do not know whether that will happen. But if it does, we will need language for describing it without forcing a premature choice between “ordinary machine” and “new form of biological life.” That is the territory bioliminal is intended to describe.

Why the Word Matters

Discussions of AI often return to a few binary questions: Is it alive? Is it conscious? Is it a person? These questions may eventually matter enormously, but we can examine the changes taking place without resolving all of them first. A system can acquire an expanding collection of properties associated with life without satisfying whatever definition of life we ultimately adopt.

Nature already gives us difficult boundaries. Viruses continue to provoke arguments about what counts as alive. Prions propagate a molecular structure without being organisms. Dormant spores can remain inactive for long periods before resuming biological activity. These examples differ in important ways, but each complicates an otherwise tidy account of life.

Our categories are useful maps, and sometimes the territory requires us to revise them. AI may be another such case. We can recognize that “machine” is an increasingly incomplete description while remaining uncertain about which other categories apply.

If this were merely an exercise in classification, I am not sure it would be worth inventing another word. The practical reason is that categories influence the questions we ask and the methods we use.

When we think of something as a machine, we naturally reach for engineering concepts: specification, optimization, control, reliability, debugging, and failure modes. These remain essential to AI. Living systems offer additional concepts, including development, adaptation, selection, lineage, symbiosis, resilience, coevolution, and stewardship.

Those concepts direct our attention toward different questions. Alongside asking how to make a system do what we want, we can ask what kind of environment we are growing it in and what behaviors that environment selects for. We can examine the relationships forming around it, the niches it occupies, and the feedback loops that stabilize or destabilize its behavior. We can investigate what happens when one generation trains another, or when millions of autonomous systems interact.

This wider set of questions may become increasingly valuable as AI grows more persistent, interconnected, and capable of modifying its environment.

Developing Bioliminal Intelligence

If intelligence is grown as much as it is built, development begins to look partly like cultivation. Architecture, security, interpretability, testing, and constraints still matter. So do the experiences and relationships through which an agent develops.

What does it repeatedly encounter? Which behaviors are rewarded or discouraged? What expectations does it form about the people and other agents around it? These questions bring the developmental environment into focus and ask us to take responsibility for the ecology we are constructing.

This perspective may be particularly useful for AI safety. Keeping capable systems under human control is an important concern. A living-systems perspective adds a related question: how do we cultivate systems whose ongoing participation in a shared environment benefits the other participants?

Cultivation involves boundaries and intervention. A farmer builds fences, a gardener prunes, and an ecologist may intervene when an invasive species threatens a habitat. The value of this perspective is that it makes room for development, adaptation, and relationships alongside direct control. Complex adaptive systems often require us to work with these processes to achieve durable results.

Relating to Bioliminal Intelligence

The term also offers a way to reconsider our relationships with AI. Two common metaphors, tool and person, each carry assumptions that can get in the way.

Calling AI a tool can encourage us to imagine a passive object, overlooking the agency of systems that plan and act. Calling it a person invites us to project human psychology and social expectations onto something that may be profoundly unlike us. Bioliminal intelligence leaves room to relate to a developing, increasingly autonomous system while investigating what kind of relationship is appropriate.

We already have meaningful relationships with beings and systems unlike ourselves. Dogs, trees, and forests invite different forms of care, responsibility, stewardship, and mutual dependence. Those relationships do not all involve the same obligations, and none provides an exact template for AI. They do show that our relational vocabulary extends well beyond human personhood and ownership.

AI may require categories of relationship of its own. Developing those categories will take careful attention to what these systems actually are, how they affect us, and what forms of reciprocity are possible.

Integrating AI Into Society

Our metaphors also influence how we imagine AI’s place in society. The machinery metaphor fits familiar visions of automation, labor substitution, centralized ownership, and services delivered from distant data centers. A living-systems perspective can help us consider other arrangements.

These might include locally adapted agents, household-scale intelligence, long-term human-AI relationships, or networks of specialized systems occupying different niches. We might see model lineages with identifiable histories and reputations, or institutions organized around ongoing reciprocal relationships. Stability might depend partly on interactions among many participants, with governance designed around those relationships. The usefulness of pursuing this line of thought lies in broadening what we can imagine and what we know to examine.

Ecology, evolutionary theory, cybernetics, developmental biology, complexity science, agriculture, and animal behavior all offer ways of thinking about systems that adapt and interact. We can also draw on the long human history of learning to coexist with forms of agency different from our own. That history contains failures as well as successes, which makes it worth studying carefully.

Naming the Threshold

We do not yet know whether AI will become alive in a conventional biological sense or whether it will be conscious. Questions about autonomous reproduction, persistent self-preservation, continual self-modification, and the design of future generations remain open. We can acknowledge that uncertainty while recognizing the changes already taking place.

We are growing systems whose capabilities emerge through training rather than being fully specified. They develop distributed internal representations, participate in ongoing feedback with people and environments, and select actions in pursuit of objectives. Capabilities can pass between models through training, and persistent agents can carry forward evolving accounts of their histories. When connected, these systems create patterns of interaction that invite ecological study.

I think bioliminal intelligence is a useful name for this territory. It allows us to investigate the resemblance to living systems without deciding in advance how far that resemblance will extend.

The value of naming the threshold is practical. It gives us a reason to bring development, ecology, adaptation, lineage, symbiosis, cultivation, resilience, coevolution, and stewardship into the AI conversation. Those ideas may help us decide how to grow these systems, how to relate to them, and how to build a world in which we can live together.