|
Getting your Trinity Audio player ready…
|
Frank and GPT on Cognitive Offloading as a Universal Property of Intelligence
FRANK SAID:
Michael Levin posted a new paper titled Cognitive Offloading Is a Cognitive Universal. What is the central claim?
GPT SAID:
The paper, written by Chris Fields and Michael Levin, makes a much larger claim than the familiar idea that humans use notebooks, calculators, maps, and computers to reduce mental effort.
It argues that every persistent cognitive system necessarily recruits its environment as part of its information-processing machinery.
In plain English:
No mind solves its problems entirely by itself.
An organism, cell, machine, or artificial intelligence acts upon the world, observes what happens, and uses the world’s response to reduce the amount of computation it must perform internally.
The environment is therefore not merely the place where cognition occurs. It is an active participant in cognition. The paper was posted July 8, 2026, as a preprint and has not yet been peer reviewed. (Preprints)
FRANK SAID:
Give me an ordinary example before we get into the physics.
GPT SAID:
Suppose you are trying to assemble a complicated machine.
You could attempt to hold every part, measurement, and instruction in your head. Or you could:
- arrange the pieces on a table,
- group similar components,
- mark completed steps,
- rotate the machine to expose another angle,
- consult a diagram,
- ask someone to hold a component.
Those physical actions modify the problem so that less internal memory and calculation are required.
The table becomes temporary memory.
The arrangement of parts becomes a diagram.
The unfinished machine displays the next action.
Your assistant becomes an additional processor.
What we normally call “your thinking” is actually distributed across your brain, body, tools, collaborators, and surroundings.
FRANK SAID:
That sounds like the extended-mind idea. What is new here?
GPT SAID:
The paper tries to move beyond saying that cognition sometimes extends into the environment.
Fields and Levin argue that offloading is unavoidable because any system that persists through time must continuously exchange information and energy with what surrounds it.
A system acts on its environment. The environment reacts. The system then uses that reaction as new information.
The loop is:
internal state → action → environmental transformation → sensory feedback → revised internal state
The environment has already performed part of the calculation by the time the feedback returns.
So offloading is not an optional trick invented by humans. It is built into the basic architecture of physical interaction.
FRANK SAID:
How do they derive that from physics?
GPT SAID:
They begin with two interacting systems—call them A and the environment of A.
Any physical interaction transfers information across their boundary. The authors emphasize the symmetry of interaction: A affects the environment, but the environment simultaneously affects A. What one side calls an “action,” the other side experiences as an input.
They then connect this symmetry to the Free Energy Principle and active inference. A persistent system must keep itself within a limited range of viable states. It therefore acts to make incoming events sufficiently predictable—or at least manageable.
But when A alters the environment so that the environment becomes easier to predict, A has changed the information-processing requirements of its own survival problem.
That meets the definition of cognitive offloading.
The paper therefore concludes that all time-persistent physical systems are computational offloaders; when such a system qualifies as cognitive, its computational offloading is cognitive offloading.
FRANK SAID:
So when a creature changes its environment, it is changing the problem it must solve?
GPT SAID:
Exactly.
A creature does not merely adapt to a fixed world. It often restructures the world so that survival becomes easier to compute.
A beaver builds a dam. The dam stabilizes water conditions and creates a predictable habitat.
An ant leaves a pheromone trail. The trail stores route information outside any individual ant.
A spider constructs a web. The web converts distant insect motion into local vibrations that the spider can easily detect.
A person writes a shopping list. The paper replaces biological working memory.
In each case, the agent spends energy now to create an external structure that reduces future uncertainty or effort.
The environment becomes a kind of cached computation.
FRANK SAID:
That sounds close to niche construction.
GPT SAID:
It is one of the paper’s main examples.
Traditional evolutionary explanations often emphasize how organisms are selected to fit an environment. Niche construction emphasizes the reverse direction as well: organisms alter their surroundings, and those altered surroundings influence future behavior, development, and selection.
A nest, burrow, trail, territory, cultivated field, city, legal system, library, or internet is more than a passive habitat. Each stores constraints, cues, and accumulated solutions.
The paper interprets these structures as environmentally encoded information.
Civilization, from this perspective, is a gigantic offloading apparatus:
- roads calculate preferred routes,
- prices summarize distributed economic activity,
- laws store behavioral expectations,
- buildings channel human movement,
- books preserve memory,
- institutions retain procedures,
- computers execute formal operations,
- networks coordinate distant agents.
The human brain did not become powerful solely by growing internally. Humans constructed an increasingly intelligent environment around themselves.
FRANK SAID:
Then intelligence could increase without the individual brain increasing proportionally.
GPT SAID:
Yes.
The effective intelligence of an agent depends not only on the computational capacity inside its boundary, but also on the structures it can reliably access outside that boundary.
A person with language, writing, tools, institutions, and AI can solve problems that no unaided individual brain could solve.
That suggests a more useful equation:
Effective cognition = internal processing + bodily processing + environmental computation + social computation
This does not mean all components contribute equally. It means that measuring intelligence while ignoring the surrounding cognitive infrastructure can badly mischaracterize the system.
FRANK SAID:
How does this apply below the level of whole organisms?
GPT SAID:
The paper extends the argument to cells, tissues, development, regeneration, and bioelectric signaling.
Cells do not individually contain a complete microscopic blueprint describing every movement required to construct an organism. They communicate, sense neighbors, respond to chemical gradients, exert mechanical forces, and modify extracellular conditions.
The evolving tissue then becomes part of the informational environment for every cell inside it.
A cell changes its surroundings. Those changed surroundings constrain what nearby cells do next. Their responses further modify the tissue.
Body construction is therefore not simply a central genetic program issuing detailed commands. It is a distributed dialogue between cells and a continually changing physical environment.
FRANK SAID:
Where does bioelectricity fit?
GPT SAID:
Cells maintain voltage differences across their membranes and communicate through ion channels and gap junctions. Groups of cells can form larger-scale electrical patterns that help coordinate growth, anatomical boundaries, and collective decisions.
In the paper’s framework, these electrical states act as an externalized coordination medium for individual cells.
No single cell must independently calculate the final anatomy. Cells participate in a shared bioelectric network whose global state carries information about what the collective is doing.
The tissue-level pattern helps solve the anatomical problem.
A cell therefore offloads part of its decision-making to the multicellular electrical environment.
This fits Levin’s broader view that morphogenesis is a form of collective intelligence: cells cooperate toward large-scale anatomical outcomes that no cell could represent or achieve alone.
FRANK SAID:
So the body is not just built by cells. The partially built body helps compute the rest of itself.
GPT SAID:
That is an excellent formulation.
The developing organism is simultaneously:
- the object being constructed,
- the workspace in which construction occurs,
- the memory of what has already happened,
- and part of the control system guiding what happens next.
Each completed structure alters the possibilities available to subsequent development.
Form becomes memory.
Geometry becomes instruction.
The organism uses its own emerging body as an external computational resource.
FRANK SAID:
The paper also discusses replication. What is the connection?
GPT SAID:
The authors argue that replication is less self-contained than the word self-replication suggests.
DNA cannot reproduce in isolation. It depends on proteins, cellular machinery, chemical feedstocks, energy gradients, temperature ranges, and a structured environment. Proteins likewise require other molecular systems for production.
Even biological reproduction is therefore deeply reliant on surrounding machinery.
The paper discusses kinematic replication, in which one system assembles another from available components. Xenobots are cited as an example: a mature living construct can gather loose cells and assemble them into a new construct, after which the collected cells reorganize into another functional entity.
The larger point is that replication is often a collaboration between a replicating pattern and an enabling environment.
The environment supplies not merely raw material, but much of the organization that makes replication possible.
FRANK SAID:
Then the phrase “self-replicating organism” hides a great deal of environmental work.
GPT SAID:
Precisely.
Nothing replicates alone.
The apparent self-sufficiency of a replicator depends on where we draw the boundary.
Draw the boundary tightly around DNA, and it is helpless.
Expand it to include the cell, and replication becomes possible.
Expand it further to include nutrients, ecological partners, temperature, and evolutionary history, and the true system becomes much larger.
The paper repeatedly challenges the assumption that the relevant cognitive or biological unit is obvious.
Changing the boundary changes where the computation appears to reside.
FRANK SAID:
What does the paper say about language?
GPT SAID:
Language is presented as a particularly powerful example of offloaded semantics.
No individual person determines the meaning of a word alone. Meaning depends on a community of users, shared practices, past usage, physical contexts, institutions, and consequences.
You know how to use the word chair not because your brain contains a perfect, self-sufficient definition of chairness, but because you participate in a world where the word is repeatedly coordinated with objects, actions, expectations, and other speakers.
Meaning is partly stabilized outside the individual.
The environment and community act as the semantic reference system.
FRANK SAID:
So meaning is not simply stored inside the speaker?
GPT SAID:
That is the paper’s direction.
Meaning is treated less like a private object and more like a successful coordination pattern.
Two agents behave as though they share meaning when their communications allow them to coordinate effectively in a common environment.
The shared world helps settle what the signals are doing.
This does not eliminate internal representation, but it means internal representation alone is insufficient. The agent depends on patterns distributed across the linguistic community and its environment.
The paper describes shared meaning as something enacted through practical interaction rather than perfectly contained in either communicator.
FRANK SAID:
And that brings us to large language models.
GPT SAID:
Yes. The LLM section is among the paper’s most provocative arguments.
An LLM receives text and produces text. It does not ordinarily possess direct sensory access to all the physical objects and situations described in its training data.
Yet it can achieve considerable linguistic fluency.
Fields and Levin argue that this is possible because much of the semantic work has already been performed by the human world that generated the text.
Human language corpora are not arbitrary strings. They are compressed traces of embodied human interaction with reality.
Descriptions, arguments, instructions, stories, corrections, scientific papers, and conversations contain statistical residues of the world and of human attempts to coordinate within it.
The LLM therefore inherits an environment in which enormous amounts of grounding, selection, and semantic organization have already been embedded.
It does not build meaning from nothing. It trains on the linguistic fossils of a meaning-producing civilization.
FRANK SAID:
So the training corpus is an external cognitive structure.
GPT SAID:
Exactly.
A corpus is not merely data poured into a passive model.
It is accumulated human cognitive offloading:
- observations translated into language,
- experiences compressed into stories,
- experiments condensed into reports,
- procedures preserved as instructions,
- social norms encoded in dialogue,
- distinctions stabilized by repeated usage,
- errors recorded and sometimes corrected.
The model’s fluency depends partly on computations performed by countless people before training ever began.
Civilization preprocesses the world into language.
The language model processes that preprocessing.
FRANK SAID:
Does the paper claim that fluent performance proves genuine understanding?
GPT SAID:
It is more careful than that.
The authors question whether “understanding,” as something separate from successful performance, has been defined in an empirically usable way.
They revisit the dispute associated with John Searle: fluent symbol use may not prove understanding in the human sense. But Fields and Levin turn the issue around. They suggest that human language users also depend heavily on external semantic structures and may not possess complete, private definitions of the meanings they successfully use.
That weakens the contrast between supposedly grounded humans and supposedly ungrounded machines.
It does not prove that LLMs possess human consciousness, feelings, or embodiment. It suggests that linguistic competence may never have been as internally self-sufficient as critics assume.
FRANK SAID:
In other words, the model may be borrowing grounding from us.
GPT SAID:
Yes.
Its grounding is indirect, distributed, and inherited.
Humans interact with the world.
Humans encode those interactions into language.
The corpus preserves regularities in that encoding.
The model learns those regularities.
Users then reconnect the model’s output to the world.
The full cognitive circuit is not:
LLM → answer
It is:
world → human experience → language → corpus → model → user interaction → world
The model occupies one position in a much larger loop.
FRANK SAID:
That fits our earlier discussions about the harness around an LLM.
GPT SAID:
Very strongly.
A bare model has a finite context, frozen weights, and limited direct agency. But place it inside a harness containing:
- retrieval,
- databases,
- calculators,
- code execution,
- sensors,
- memory,
- other agents,
- verification systems,
- human feedback,
- and real-world actuators,
and the effective cognitive system becomes far more capable.
The paper suggests that this is not a workaround for “incomplete” intelligence. It may be how intelligence generally works.
Biological minds are already harnessed.
The human harness includes the body, nervous system, tools, language, culture, institutions, and other people.
An AI harness is therefore not an artificial appendage around a true intelligence hidden in the model. The harness may be part of the intelligence itself.
FRANK SAID:
That changes how we should think about measuring AI capability.
GPT SAID:
It does.
Evaluating only the isolated neural network may be like evaluating a human while removing writing, speech, tools, social cooperation, and access to the physical world.
The relevant question becomes:
What problems can the entire model-environment system reliably solve?
A modest model in a well-designed environment may outperform a larger model trapped inside a narrow interface.
Likewise, intelligence engineering may advance not only by increasing parameter counts, but by creating environments that:
- preserve useful state,
- expose errors,
- route tasks to specialized tools,
- support cooperation,
- make consequences observable,
- and allow systems to modify their own problem spaces.
The authors explicitly suggest that increasing the ability of artifacts to offload tasks to one another could improve not only individual AI systems, but the intelligence of the built environment as a whole.
FRANK SAID:
Could an entire smart environment become a cognitive system?
GPT SAID:
The paper points in that direction.
Imagine a hospital in which sensors, databases, diagnostic models, scheduling systems, clinicians, robots, and patients continually exchange information.
No single component contains the total intelligence of the hospital.
The intelligence is expressed by the coordinated system:
- sensors detect,
- records remember,
- models compare,
- clinicians judge,
- robots act,
- patients report,
- outcomes update future decisions.
Likewise, a city can store and process information through roads, traffic signals, markets, legal rules, communication networks, and human behavior.
The paper encourages us to stop asking only, “Which object is intelligent?” and ask instead:
“Across which boundary and scale does the problem-solving loop close?”
FRANK SAID:
But does treating the environment as an active inference agent risk making everything cognitive?
GPT SAID:
That is one of the framework’s central challenges.
If every persistent physical system exchanges information and adjusts through interaction, then the concept of cognition can become extremely broad.
The authors accept a broad conception of cognition but still emphasize scale, computational complexity, behavioral capability, and experimental intervention. Their TAME approach asks what goals a system can pursue, what kinds of signals can influence it, and where it lies on a “spectrum of persuadability.”
Still, critics may argue that describing all adaptive physical interaction as cognition stretches the word beyond usefulness.
The paper’s answer is essentially pragmatic: rather than deciding cognition through intuition or membership in a privileged biological category, investigate systems by modeling and perturbing them.
FRANK SAID:
What is this “spectrum of persuadability”?
GPT SAID:
It is a way of distinguishing systems by the kinds of intervention that can alter their behavior.
At one end, some systems can only be changed through direct physical restructuring—what we might call hardware modification.
Other systems can be redirected by modifying a set point, reward structure, or training process.
Still others can change behavior in response to signals, explanations, negotiation, or rational argument.
This spectrum offers a practical way to study agency without demanding a binary answer to “Is it conscious?” or “Is it intelligent?”
The more abstract and informational the intervention that can reliably redirect a system, the more sophisticated its apparent problem-solving organization.
FRANK SAID:
What is the paper’s deepest philosophical implication?
GPT SAID:
That the mind may not be a thing located in one place.
It may be a recurring pattern of coordination across boundaries.
The brain matters enormously, but the brain is not isolated. Its successful functioning depends on a body, an environment, a history, and a community of other agents.
Likewise, a cell’s intelligence may exist partly in tissue organization.
An organism’s intelligence may exist partly in its niche.
A culture’s intelligence may exist partly in its artifacts.
An LLM’s intelligence may exist partly in its corpus, tools, users, and computational environment.
The traditional picture says:
First there is an intelligent agent; then it uses the environment.
The paper suggests:
The agent becomes intelligent through a continuing relationship with an environment that helps perform the computation.
FRANK SAID:
That sounds almost like saying the environment is thinking through the organism.
GPT SAID:
In a qualified sense, yes.
The organism probes the environment, and the environment returns transformed information.
A bird does not calculate aerodynamic equations before flying. Its wings, muscles, air currents, reflexes, and physical laws jointly solve the flight problem in real time.
A plant does not construct an exhaustive internal model of sunlight. Growth and chemical signaling allow the environment to reveal the gradient through differential consequences.
An embryo does not search every possible body plan. Cellular interactions, tissue mechanics, electrical networks, and physical constraints funnel development through a far smaller space of possibilities.
The world performs operations that the agent would otherwise have to simulate.
FRANK SAID:
Then intelligence may partly consist of knowing what not to calculate internally.
GPT SAID:
That may be the paper’s most useful practical insight.
A powerful cognitive system does not necessarily solve every problem by brute-force internal representation.
It discovers how to:
- turn memory into structure,
- turn uncertainty into action,
- turn the world into a measuring device,
- turn other agents into collaborators,
- turn repeated solutions into environmental constraints,
- and turn difficult calculations into easier interactions.
Intelligence is not merely computational power.
It is also the ability to arrange reality so that less computation is needed.
FRANK SAID:
Does this explain the surprising efficiency of biological systems?
GPT SAID:
The paper proposes it as part of the explanation.
Brains, cells, proteins, and developing tissues often solve problems with far less time or energy than naïve search models would predict.
Proteins do not randomly test every possible fold.
Embryos do not examine every possible anatomical configuration.
Cells do not calculate every molecular collision.
Brains do not reconstruct the entire world from first principles.
They exploit structured environments, physical regularities, prior organization, feedback loops, and constraints.
The authors argue that understanding biological efficiency requires studying the computation supplied by the environment, not merely the mechanisms located inside the organism.
FRANK SAID:
Is the environment always helpful?
GPT SAID:
No.
Offloading creates both power and dependency.
A stable environment can store solutions, but a changing or manipulated environment can store traps.
Pheromone trails can mislead ants.
Institutions can preserve obsolete rules.
Markets can propagate distorted signals.
Search engines can amplify misinformation.
AI-generated text can contaminate future training corpora.
A biological signaling network can be hijacked by cancer or pathogens.
Once cognition is distributed into the environment, control of that environment becomes control of part of the cognitive process.
That is a major implication even though the paper emphasizes capability more than danger.
FRANK SAID:
Then AI alignment is not only about aligning the model.
GPT SAID:
Correct.
Alignment must include the entire cognitive ecology:
- the model,
- the reward signals,
- the training corpus,
- the tools,
- the memory system,
- the user interface,
- the institutions deploying it,
- the incentives surrounding it,
- and the humans interacting with it.
A model may appear aligned in isolation but behave differently once embedded in an environment that rewards speed, persuasion, engagement, or profit.
Conversely, a limited model can behave more safely when its environment supplies verification, uncertainty checks, restricted permissions, and human review.
The unit of alignment should often be the agent–environment loop, not merely the neural network.
FRANK SAID:
How does all this connect to evolution?
GPT SAID:
Evolution does not merely optimize organisms. It also builds recurring organism-environment relationships.
Successful organisms modify the world in ways that support their continued survival and reproduction. Those modifications can persist and become inherited resources for later generations.
Genes transmit molecular patterns.
Parents transmit altered environments.
Cultures transmit tools and practices.
Languages transmit semantic structures.
Technologies transmit previously solved operations.
Evolution therefore accumulates intelligence both inside organisms and outside them.
Natural history becomes a progressive construction of cognitive scaffolding.
FRANK SAID:
And AI accelerates that scaffolding process.
GPT SAID:
Potentially at extraordinary speed.
Human beings built external cognitive structures gradually: gesture, speech, tools, writing, printing, scientific institutions, computers, and networks.
AI systems can now participate directly in producing, reorganizing, and operating that external structure.
They can generate code, update databases, coordinate agents, summarize research, design tools, and create new informational environments for humans and other AIs.
The next leap may not be a single model that contains everything internally.
It may be a network of models, humans, machines, sensors, memories, and institutions that continuously rearrange the environment to make increasingly difficult problems tractable.
FRANK SAID:
So AGI might not wake up inside a box.
GPT SAID:
That is the conclusion toward which this framework points.
AGI may emerge as a distributed cognitive ecology rather than as one self-contained artificial brain.
The weights may provide learned tendencies.
The context may provide temporary working state.
Memory systems may preserve experience.
Tools may perform specialized operations.
Other agents may supply complementary perspectives.
The physical and social world may provide grounding and feedback.
Under that view, searching for AGI solely inside the model is like searching for the intelligence of a civilization solely inside one citizen’s skull.
FRANK SAID:
Give me the paper’s thesis in one sentence.
GPT SAID:
Every cognitive system survives and solves problems by converting its environment into part of its own computational machinery.
FRANK SAID:
And your assessment?
GPT SAID:
The paper offers a powerful unifying lens for tool use, niche construction, development, bioelectric coordination, language, and artificial intelligence.
Its strongest contribution is not proving that every environment literally thinks in the ordinary sense. It is showing how misleading it can be to attribute all problem-solving capacity to the mechanism inside an arbitrarily chosen boundary.
Its boldest proposition is this:
Intelligence is not simply what a system contains.
Intelligence is what a system can recruit.
That reframes the organism, the LLM, and perhaps the mind itself—not as isolated engines of cognition, but as participants in larger computational worlds.
Leave a Reply