The case for intelligence as learned transduction

What if intelligence is better understood as a transformation than as a substance? Token Processing Intelligence Theory, or TPIT, starts from that proposition. A “token” is simply an operational difference: something a system can distinguish and use. An adaptive structure transforms those differences into further differences that affect what happens next.

The Learned Value Transducer, or LVT, is the control-oriented form of the same idea. Measurements enter an adaptive system and are converted into values that regulate a process. Those values need not be explicit numbers. They can be neural signals, chemical states, priorities, expectations or any other internal difference that changes subsequent behaviour.

A model, not a metaphor

The comparison with large language models becomes useful only if it is made carefully. Humans are not literally transformer networks, and they are not copies of a single architecture trained on different personal datasets. The stronger analogy is that every person is a custom learned system. Inherited biology, sensory capacities, memory, developmental plasticity, reward systems and learning dynamics can all differ before individual experience is added.

Experience then changes the system further. Two people exposed to the same event need not receive the same effective input: they may notice different variables, compress them differently, assign different values to possible outcomes and possess different actions with which to respond. Their actions alter the environment, which changes the next input. The learner therefore helps generate its own future training data.

This matters because it turns individual variation from an inconvenience into an expectation. Skill is not just accumulated information. It is the result of information interacting with a particular learning system, at a particular developmental state, through a particular history of feedback.

Where the complexity comes from

The more consequential claim is about emergence. Complexity does not have to reside primarily inside each component. It can arise from the way components are connected. One transducer produces an output that becomes another transducer’s measurement; the second changes a process that alters the first transducer’s future conditions. Add many such units, recurrent loops, delays, nonlinear responses and changing connection strengths, and the network acquires behaviour that cannot be inferred by inspecting any one node.

The pattern can recur across scales. Molecular regulatory systems participate in cells; cells in organs; neural circuits in people; people in firms, markets and cultures. At each level, lower-level transformations can combine into a higher-level transformation. The higher level does not have to reproduce every microscopic interaction beneath it to have stable, observable behaviour of its own.

This is also an answer to the prediction problem. A system containing millions or billions of interacting LVTs may be impossible to model bottom-up in useful detail. But a second learned system can observe the aggregate system and learn an effective mapping between measurements and later outcomes. Prediction becomes another instance of learned transduction.

Why large language models matter

Large language models are a striking example of this strategy. Their training material contains traces of an enormous human network: explanation, argument, code, convention, science and culture produced by people interacting over generations. An LLM does not reconstruct every person who contributed to those traces. It learns regularities in the collective output.

That does not make an LLM a complete model of society. It shows something narrower and more important: a learned system can capture useful structure generated by an emergent system without possessing a microscopic theory of all its parts. In TPIT terms, one learned transducer can approximate aspects of the higher-level transformation produced by a network of other learned transducers.

Once models are returned to society, the recursion deepens. Human outputs train models; model outputs become human inputs; people change their behaviour; institutions adapt; later models are trained on a world partly shaped by earlier models. The relevant system is no longer simply “human” or “AI”, but the coupled network they form.

What the experiments show

Action changes learning. In the 1963 kitten-carousel experiment, Richard Held and Alan Hein gave paired kittens closely matched visual exposure while allowing only one kitten to control its own movement. The actively moving animals developed normal visually guided behaviour; the passive animals did not. The informational content of exposure was not enough. The action–consequence–feedback loop mattered.

Signals can acquire new functions. In Paul Bach-y-Rita’s sensory-substitution experiments, camera information was converted into tactile stimulation and blind participants learned to use those patterns for spatial discrimination. A signal’s functional role was not fixed by its physical channel; training established how it could be used.

Training changes the learner. Studies of London taxi drivers found experience-related differences in hippocampal structure, while controlled juggling studies found measurable structural changes following training. Human learning is therefore not cleanly separable into fixed architecture plus changing content. Experience can alter the machinery that processes later experience.

Different learners respond differently to similar data. Twin studies show substantial inherited contributions to psychological variation. Critical-period research shows that similar linguistic input can have very different effects at different developmental stages. Deliberate-practice research finds meaningful but incomplete relationships between practice and performance. Working-memory training tends to transfer strongly to nearby tasks but weakly to distant ones. Together, these findings fit a population of differently specified, differently plastic models better than a population of identical blank learners.

Learned value can regulate behaviour. Work by Wolfram Schultz, Peter Dayan and Read Montague linked changes in primate dopamine activity to reward-prediction error. The functional pattern is close to the LVT idea: measurements of a situation are transformed into learned value signals that influence future action and learning.

Adaptive transformation does not require a brain. Slime mould has shown stimulus-specific habituation to normally aversive substances, including recovery after the stimulus is removed. Single-celled Stentor has shown history-dependent changes in responsiveness. These organisms are not miniature humans, but they demonstrate that elementary learning-like transformation can be implemented in very different substrates. Claims about plant learning are more disputed and deserve a higher evidential bar.

The strongest test is at the network level

Experiments on groups make the emergence claim more concrete. Anita Woolley and colleagues found evidence for a group-level collective-intelligence factor that was not reducible to the intelligence of the single strongest member and was associated with interactional properties such as social sensitivity and conversational turn-taking.

Winter Mason and Duncan Watts showed that communication networks could alter how groups explored a difficult search problem and could improve collective performance relative to isolated problem-solvers. Matthew Salganik, Peter Dodds and Watts found an even sharper feedback effect in an artificial music market: when participants could see previous participants’ choices, outcomes became more unequal and less predictable as social influence increased. The songs had not changed. The information flowing between people had.

These findings do not imply that network structure always dominates outcomes. Other experiments have found weak or absent topology effects under particular tasks and incentives. That is exactly the constraint a serious account needs. Emergence depends on what the nodes transform, what the links transmit, how feedback operates and what process is being controlled.

What follows

TPIT and LVT therefore suggest a shift in the unit of analysis. Intelligence can be treated as learned transformation within systems, while complexity can arise from recurrent coupling among them. Individual humans are custom learned models embedded in networks of other custom models. Those networks generate higher-order behaviour, and further learned systems can model that behaviour at the level at which it becomes observable.

The implication for AI is broader than the claim that machines can imitate people. Large language models already show that a learned system can absorb the traces of collective human activity and recover useful predictive structure without a complete bottom-up theory of the society that produced those traces. As such models are connected back into human institutions, they become components of the very networks they model.

The most interesting object may therefore be neither the biological individual nor the artificial model. It may be the recursively connected system in which learned transducers observe, influence and train one another across multiple scales.