The argument begins by treating intelligence functionally. TPIT supplies the broad unit: an operational difference is transformed by learned or adaptive structure into another difference. LVT narrows the idea to control: measurements are transformed into values that influence a process and return through feedback. This creates a common vocabulary for systems that may differ radically in physical implementation.
The human–LLM comparison is then used as a simplifying device, not as a claim of literal equivalence. A person is treated as a custom learned model whose architecture, initial conditions, plasticity and experience can all differ from those of another person. That formulation makes individual variation part of the theory rather than something to be explained away, and it accommodates evidence from genetics, development, expertise and neuroplasticity.
The central move is to place emergence in the coupling among systems. Outputs become other systems’ inputs; those responses alter later conditions; feedback repeats. Once many transducers are linked in this way, network-level behaviour can become qualitatively different from the behaviour of any single component. This gives the framework a recursive form: lower-level transformations can combine into higher-level transformations across cells, brains, people, organisations and societies.
Prediction follows from the same recursion. A higher-level system does not have to calculate every interaction inside a complex network. It can learn the effective relationship between observable measurements and later outcomes. Large language models provide the contemporary example: they learn regularities from the accumulated outputs of a vast human network without reconstructing the internal state of every person who contributed to those outputs.
The evidence was selected to test each part of that structure. Held and Hein isolate the importance of action–feedback coupling. Bach-y-Rita shows that a signal can acquire a new operational role through learning. Neuroplasticity studies show that training can modify the learner itself. Twin, critical-period, practice and transfer research establish that learners differ in architecture, state and response to experience. Dopamine research provides a concrete example of learned value signals involved in control. Slime mould and Stentor extend adaptive transformation beyond nervous systems. Group and social-influence experiments show that changing connections and feedback among agents can change collective outcomes.
The resulting essay is deliberately narrower than a claim that all adaptive systems are intelligent. Its strongest empirical footing lies in learned transformation, feedback, plasticity, value-guided control and network effects. Substrate independence is treated as a hypothesis about implementation, not as permission to erase distinctions between organisms, brains and machines. Where evidence is contested, as in some claims about plant learning, the article says so.
The synthesis can be reduced to one chain. Differently specified systems learn from different histories. Their actions help generate the data that train them next. Their outputs become inputs to other systems. Recurrent coupling produces higher-order behaviour. Higher-order learned systems can then model those emergent transformations directly. Large language models are important because they make the final step visible: a learned model can recover useful structure from the traces of a complex human network and can then be connected back into that network.