A research perspective on Norbert Wiener, NeuroAI, and why the unit of intelligence may be larger than the model
The more I think about NeuroAI, the less satisfied I am with one of its most natural questions:
Which artificial model best resembles the brain?
It is a good question. Neural predictivity, representational similarity, and artificial networks as models of neural systems have made brain–machine comparison quantitative in a way that earlier theories rarely could. Frameworks such as Brain-Score are valuable precisely because they force intuitions about brain-likeness to survive contact with data (Schrimpf et al., 2018).
But I increasingly think the question is incomplete.
Much of the comparison still begins with something like:
stimulus → brain representation
versus
stimulus → model representation
and asks how similar the two endpoints are.
What keeps bothering me is everything left outside the arrows.
Brains do not normally receive a fixed dataset, produce a representation, and stop. They sit inside bodies. Bodies act. Actions change what is sensed next. Experience changes the system that generated the action. Other agents respond. The data-generating process moves because the intelligent system is inside it.
This is where I find Norbert Wiener unexpectedly useful.
Not because Wiener had a theory of modern NeuroAI waiting to be rediscovered, and not because every later idea about prediction, embodiment, or active inference is secretly cybernetics. I use Wiener more narrowly: as a lens for choosing the unit of analysis.
Read together, Behavior, Purpose and Teleology (Rosenblueth, Wiener, & Bigelow, 1943), Cybernetics: Or Control and Communication in the Animal and the Machine (Wiener, 1948), and The Human Use of Human Beings (Wiener, 1950/1954) trace a striking expansion of the system boundary:
environment → sensing → action → environment
then
experience → learning → changed controller → changed experience
and eventually
human ↔ machine ↔ institution ↔ society
That progression has changed how I think about NeuroAI. Wiener offers less a particular model of intelligence than a different way of deciding what the scientific object should be.
For a Wienerian NeuroAI, a natural unit of analysis is the adaptive organism–environment loop.
In The System Is the Model, I frame the same issue as a boundary question: at what scale does the explanatory loop close? Here I want to make that question operational. If the relevant unit is a loop, then brain–model comparison should eventually compare what happens around the loop, not only what is represented at one point inside it.
Once I take that seriously, representation, prediction, action, embodiment, learning, robustness, and even parts of alignment start to look like different views of the same dynamical problem.
There is a second lesson that appears only after following the loop far enough outward. A control system may become extraordinarily good at pursuing an objective without becoming any wiser about whether that objective should have been chosen.
So this essay is about two ideas that I now think belong together:
intelligence is larger than the model, and optimization is smaller than judgment.
The first shift: from output to purpose
The place to begin is the 1943 paper Behavior, Purpose and Teleology, written by Arturo Rosenblueth, Norbert Wiener, and Julian Bigelow.
What I find most useful in that paper is methodological. Before explaining what a system is made of internally, the authors propose that we can characterize how its behavior depends on its surroundings. An animal and a machine can therefore be compared without pretending they are materially the same.
A cat and a servomechanism are built from radically different substrates. Yet both can be described in terms of inputs, internal processes, actions, and consequences.
For NeuroAI, that separates at least two kinds of similarity:
implementation similarity
and
organizational similarity
A model may be unlike cortex in physical implementation while instantiating a control principle relevant to neural function. Conversely, two systems may produce similar representational geometry through very different causal organizations.
This loosens an assumption NeuroAI sometimes makes too quickly: that the most interesting notion of brain-likeness must be found inside the model.
Purpose without a mysterious final cause
The deeper move in the 1943 paper is its treatment of purpose.
Purpose had long carried the flavor of a final cause: behavior appears organized by a future state that does not yet exist. Rosenblueth, Wiener, and Bigelow show how to make the problem mechanical without making it trivial.
In modern notation:
error = desired state - current state
action = function(error)
The future does not need to reach backward in time. The system needs a current state, some organization around a preferred state, and a mechanism by which deviation changes action.
A thermostat is the usual minimal example. The point is not that a thermostat is intelligent. It is that goal-directedness can be implemented as closed-loop organization.
Animals make that organization vastly richer. They maintain posture, pursue food, avoid threats, reach for objects, coordinate with others, and switch strategies as the world changes. In each case, the action matters because of what happens next.
An impressive response is not yet the same thing as purposive organization.
A feedforward mapping can produce the right answer:
input → output
but a purposive system forces us to continue the diagram:
input → action → consequence → new input → new action
The output acquires its full meaning only when it is allowed to return through the world.
Prediction is interesting when it changes action
Rosenblueth, Wiener, and Bigelow also distinguish non-predictive from predictive forms of purposeful behavior. A system that reacts only to where a moving target is now may fail because action takes time; effective control may require anticipating where the target will be when the action lands.
Prediction in machine learning is often written as:
past data → prediction of future data
Prediction inside a controller looks more like:
predict future state → choose action → alter future state → observe consequence
The prediction is no longer merely evaluated against future data. It helps produce the future data.
An eye movement changes visual input. Locomotion changes visual, vestibular, and proprioceptive signals. Reaching changes both scene and body. A social action changes what another agent does next.
The future data distribution is partly endogenous to the policy.
That is why embodiment is more than an engineering add-on. It changes the statistical problem: the learner is not just fitting a distribution; it is helping create the distribution it will later have to understand.
Sometimes the failure is in the loop
The same paper uses pathological feedback to make another lasting point. In discussing intention tremor, the authors consider what happens when correction itself becomes unstable.
positive error → overcorrection → negative error → overcorrection → ...
With the wrong gain, delay, or dynamical coupling, a mechanism designed to reduce error can generate new error instead.
This changes the explanatory instinct. Instead of always asking which module is broken, we can ask whether the coupled process has become unstable.
brain area → function
may sometimes be less useful than
brain area → role in a distributed feedback process
Once the loop becomes the object, success and failure become properties of dynamics, not only properties of parts.
The second shift: from substrate to organization
Wiener’s 1948 Cybernetics turns the logic of the earlier paper into a general research program. The subtitle—Control and Communication in the Animal and the Machine—is already the argument.
The common language is not neurons or digital computation. It is control and communication.
A system that regulates its behavior needs information about the world and about the consequences of its own actions. Noise matters because corrupted information produces poor regulation. Delay matters because feedback about an old state may arrive too late. Prediction matters because action often has to compensate for a future condition. Stability matters because feedback can amplify deviation as easily as suppress it.
The minimal picture is:
world → measurement → controller → action → world
with the return path:
world → new measurement → updated action
This is why I think “the brain is a computer” is too weak a reading of cybernetics. The more interesting claim is that animals and machines may sometimes be compared at the level of organization, even when their components are radically different.
That gives me a useful hierarchy for NeuroAI. When two systems look similar, I want to know:
- Do they represent similar variables?
- Do they transform those variables through similar computations or dynamics?
- Do those states play similar causal roles inside behavior?
The third question is the easiest to lose when brain-model comparison stops at representational similarity.
Representation as a state for doing something
Modern neuroscience is extraordinarily good at measuring representation. Encoding models, representational similarity analysis, neural predictivity, and deep-network comparisons let us ask which artificial feature spaces explain variance in neural populations.
These are essential constraints. But I no longer think “what is represented?” is always the final question.
A control perspective asks:
What must be represented for the organism to do what it needs to do next?
An organism does not require a photographic reconstruction of the world. It requires internal states sufficient for effective action under uncertainty.
world → internal state → action → future world
This suggests a more functional idea of representation:
A useful representation preserves the information that matters for control.
Later control theory gives us language Wiener did not use in this form. Observability asks whether hidden states can be inferred from measurements. Controllability asks whether action can move the system through the relevant state space.
Translated back into cognition, perception can be viewed as making behaviorally relevant latent variables observable. Retinal input does not hand us three-dimensional pose, hidden causes, future trajectories, or another agent’s intention. The nervous system estimates variables that make action possible despite partial observability.
Action depends on what can be changed. A chair is not only a visual object; relative to a body it can be sittable, movable, reachable, or obstructing. Cisek’s affordance competition hypothesis is one modern neuroscientific expression of this idea: sensory information can continuously specify multiple possible actions that compete for selection rather than passing through a clean perception–decision–action pipeline (Cisek, 2007).
The brain, on this reading, is not just describing an environment. It is organizing possible relations with it.
Internal models are useful because consequences matter
Sensorimotor neuroscience gives a concrete version of the same point. Wolpert, Ghahramani, and Jordan (1995) argued for internal models of sensorimotor dynamics:
current state + candidate action → predicted next state
Prediction is useful because feedback is noisy and delayed.
Optimal feedback control makes the idea sharper. Todorov and Jordan (2002) showed how biological movement can be understood not as faithful reproduction of one desired trajectory, but as selective correction of the deviations that matter to the task.
preserve task success under noise and redundancy
rather than
reproduce the same movement exactly
The controller need not care equally about every degree of freedom. It needs to preserve what matters for the goal.
Conant and Ashby’s “good regulator theorem” formalized a related cybernetic intuition: under its assumptions, a maximally successful and simple regulator must stand in a model-like relation to the system it regulates (Conant & Ashby, 1970). I do not read that theorem as saying that every controller contains an explicit predictive world model. The narrower point is that successful regulation requires internal organization that reflects relevant structure in what is being regulated.
Predictive internal models are one realization of that principle. The broader lesson is:
Capture the regularities that matter for acting in the world, not the world in full.
Predictive coding and active inference are relatives, not synonyms
It is tempting to draw a straight historical line from Wiener to predictive coding or active inference. That would make the story cleaner and less accurate.
Rao and Ballard’s predictive coding model proposes that higher levels predict lower-level neural activity while feedforward pathways carry residual errors (Rao & Ballard, 1999). Friston’s free-energy principle and active inference go much further: perception updates beliefs to account for sensory observations, while action changes observations through intervention in the world (Friston, 2010; Friston et al., 2016).
The family resemblance to cybernetics is real: prediction, feedback, regulation, adaptation. But active inference is not Wiener rewritten in Bayesian notation.
The connection I find useful is at the level of the problem:
Wiener helps define a problem space in which intelligent systems regulate themselves through information and feedback; later theories propose different mathematical solutions inside that space.
Cybernetics becomes more useful when it is allowed to be an ancestor of questions rather than the secret origin of every answer.
The loop begins to learn
In The Human Use of Human Beings, Wiener discusses systems in which feedback can alter not only the current response but the pattern by which future responses are generated. The controller itself becomes plastic.
That suggests two nested timescales:
state_t → action_t → state_(t+1)
and
parameters_t → experience → parameters_(t+1)
The first changes what the system does now. The second changes how it will act later.
Learning is feedback acting on the organization of the controller itself.
That framing makes development difficult to treat as an afterthought. Synaptic plasticity, neuromodulation, memory, structural change, and development reshape the system that will produce the next behavior. Matching an adult endpoint is not necessarily the same as explaining how that endpoint arose.
Recent NeuroAI work makes this distinction explicit. Ayzenberg, Bonner, and Bayet (2026), for example, argue for developmental alignment: a model of human intelligence should capture developmental changes in architecture, experience, and learning objectives, rather than approximate only adult competence.
An intelligent system is not merely a controller. It is a controller that becomes a different controller through interaction.
The environment is part of the learning algorithm
Once action and learning are placed in the same loop, something else follows:
The agent partly chooses the data from which it learns.
A passive learner receives:
dataset D → model θ
A behaving agent generates a trajectory:
policy π → visited states → observations → learning → new policy π'
The policy changes the future data distribution.
Exploration, eye movements, attention, curiosity, play, and experimental action therefore help determine what will be learned. The animal is both controller and experimenter.
The same logic appears in contemporary AI. A recommender changes behavior that later becomes training data. An agent chooses tool calls that determine what information it sees next. A robot positions its sensors. Human responses to model outputs become future context or training signal.
model → policy → data → model
At that point, data generation is part of the effective dynamics of the system.
Before intelligence can optimize, an organism has to remain an organism
Wiener also places cybernetics inside a broader discussion of life, information, organization, and entropy. Living systems persist by maintaining constrained forms of organization despite disturbance and noise.
In modern language, one might imagine a restricted region of viable states:
state_t ∈ viable region
That notation is ours, not Wiener’s. The genealogy is broader: Cannon’s homeostasis emphasized maintaining physiological variables within workable ranges, while Ashby’s Design for a Brain developed adaptation in relation to stability and essential variables (Cannon, 1932; Ashby, 1952). Wiener helped place these regulatory problems inside a larger language of information, feedback, and organization.
The point matters because biological purpose is not quite like an arbitrary reward function. Before an organism solves an abstract task, it must remain an organism.
Homeostasis is not the whole of intelligence, but it is a revealing prototype. It leads to a more basic question than benchmark performance:
How does an adaptive system preserve its organization while acting in an uncertain world?
Control, learning, embodiment, development, and biological constraint now look less like separate modules than different parts of the same problem.
What this changes for NeuroAI
This is where the historical reading becomes a research perspective.
NeuroAI already has several ambitions: building AI from principles of biological intelligence, using artificial models to explain neural systems, and using AI as a tool for neuroscience (Hassabis et al., 2017; Sadeh & Clopath, 2025). What changes, for me, is the target of comparison.
The field is already moving beyond static representational matching. Feather, Khosla, Murty, and Nayebi (2025) propose a NeuroAI Turing Test combining behavioral and representational equivalence. Behavior alone underdetermines mechanism; neural similarity without behavioral adequacy is not enough either.
A Wienerian perspective suggests another step.
Suppose two systems behave similarly and exhibit similar hidden-state geometry. I still want to know whether those states occupy comparable causal positions inside the loop. Do they guide similar actions? Do comparable perturbations produce comparable corrections? Do they fail similarly under delay and noise? Does experience modify the controller in comparable ways?
The additional standard I have in mind is causal-role equivalence under closed-loop intervention: not merely whether two systems behave similarly or contain similar representations, but whether corresponding internal states are used in similar ways when the world pushes back.
This is deliberately narrower than claiming that closed-loop evaluation itself is new. Zador and colleagues (2023) proposed an embodied Turing test in which artificial animal models interact with the sensorimotor world at the level of biological counterparts. D’Angelo and colleagues (2026) evaluate controller, body, and dynamic environment together in embodied neuromorphic benchmarking. Krakauer and colleagues (2017), from another direction, argue that behavior is indispensable for understanding neural mechanism.
What I want to add to that agenda is a stronger requirement on internal mechanism. A model could satisfy a behavioral or embodied benchmark—and even match neural geometry—while the matched states occupy different causal roles in correction, sensing, or adaptation. That is exactly the kind of model I would want this framework to reject.
A Wienerian benchmark
A conventional comparison might be:
stimulus → brain activity
stimulus → model activity
followed by:
similarity(brain, model)
A NeuroAI Turing Test adds behavior. A cybernetic benchmark keeps both, then adds the missing arrows:
environment → sensing → internal state → action → new environment
The dimensions I care about are state estimation, prediction of action consequences, feedback correction, robustness to delay and perturbation, active sensing, learning and development—and representation itself, not only whether internal states look alike but whether they are used alike.
In shorthand:
behavior + representation + dynamics + perturbation response + control + adaptation
This is not an argument that every neuroscience experiment needs a robot attached to it. The claim is simpler:
A larger theory should be able to say where a representation sits in the loop.
A neural state becomes more mechanistically informative when we know not only what predicts it, but what it predicts, what perturbs it, what action depends on it, and how feedback changes it.
What would make this a research program?
If this perspective is more than a change of vocabulary, it should generate comparisons that static representational matching can fail.
Three classes of experiment seem especially diagnostic:
- Perturbation and recovery. Put biological and artificial systems into matched disturbances—sensory noise, delays, altered dynamics, missing channels—and compare not only endpoint performance but the trajectory of correction.
- Active information acquisition. Give both systems uncertainty they can reduce through action. Compare what they choose to sense, where they look, which action they use to disambiguate the state, and how those choices reshape later representations.
- Learning trajectories. Match interaction histories rather than only adult endpoints. Ask whether comparable experience produces comparable changes in policy, representation, error correction, and exploration over time.
“Matched” cannot simply mean numerically identical. A 100 ms delay is not the same intervention for systems with different bodies, control bandwidths, or behavioral timescales. The more defensible comparison is functional: delay relative to the system’s normal control-loop timescale, noise relative to baseline sensory uncertainty, or perturbation magnitude relative to the range over which the system normally corrects. Cross-system experiments need matched operating regimes, not merely matched physical units.
A model could match neural geometry impressively while diverging on all three. That would make it brain-like in one narrow and useful sense, but cybernetically unlike the organism. Conversely, a model with imperfect representational alignment might still reveal a shared control principle if its perturbation, sensing, and adaptation dynamics line up.
That is the kind of disagreement between metrics I would want a closed-loop NeuroAI program to expose rather than average away.
This is also where I want to keep the scope precise. Feedback, regulation, and adaptation are not a sufficient definition of intelligence; many narrow controllers have all three. I am making a methodological claim instead: for many NeuroAI questions, closed-loop causal organization is a more discriminating object of mechanistic comparison than representation alone. Transfer, strategy reorganization, abstraction, and goal change may require additional principles.
LLMs make the boundary visible
Large language models make this distinction easy to see.
The core pretrained objective is approximately:
context → distribution over next token
That objective can produce remarkably rich internal representations. But next-token prediction by itself is not yet a controller of a persistent external environment.
Autoregressive generation closes a small loop over the model’s own context, and dialogue creates another because model outputs change human responses. The stronger transition happens when an output changes a persistent external state that the system can later sense again:
observe → infer → act → change environment → re-observe consequence → adapt
Modern LLM systems increasingly cross that boundary through tools, memory, reinforcement learning, browsers, code execution, robots, and other agents. At that point, the scientific object is no longer the language model alone:
model + memory + tools + policy + environment + feedback
“Agentification” is therefore a change in system type, from a predictive component toward a control architecture.
The question I keep coming back to is not only:
How good is the prediction?
but:
What happens when the prediction is inserted into a loop that can act?
That question completes the computational journey—but it also exposes a missing variable. A loop can regulate, learn, and act only relative to some notion of what counts as success. Once the system becomes capable enough to shape the world, the source of that purpose becomes part of the system-level problem rather than an external detail.
This is where The Human Use of Human Beings stops looking like a separate ethical appendix and starts to look like the final turn of the same argument.
The final turn: who chooses the purpose?
Cybernetics can describe how a system regulates itself around an objective. It cannot, by itself, tell us which objective deserves to organize the system.
In modern optimization notation:
π* = argmin J(π)
Engineering asks how to find π*.
The question outside the optimization is:
Whose J?
Who chose the objective? Whose error signal counts? Which deviations count as failure? Who bears the consequences when optimization succeeds exactly as specified?
This is where Wiener becomes unexpectedly contemporary. His concern is not only that machines might replace people. Human beings themselves can be reorganized as components inside an optimizing system.
The danger is not merely:
machine replaces human
but also:
human becomes a component inside a machine-like objective
A recommender can improve engagement while changing the preferences it later claims only to predict. I follow that case from the human side in The User Is Not a Vector. A workplace optimizer can increase throughput while narrowing human judgment. An autonomous agent can satisfy a proxy while violating the intention that made the proxy useful.
The cybernetic success condition and the human success condition are not automatically the same.
In contemporary language:
optimization ≠ specification
Making the optimizer better does not tell us what should be optimized. Making the predictor more accurate does not tell us which predictions should guide action.
For NeuroAI, this matters because biological intelligence does not begin with a clean externally supplied objective. Drives, needs, development, social interaction, and value are part of the system that produces behavior. If we model only the machinery that turns goals into actions, we have explained only part of agency.
And once intelligent systems enter social environments, the loop widens again:
human ↔ machine ↔ institution ↔ society
Predictions alter decisions. Decisions alter behavior. Behavior alters data. Data alters later models.
Recommendation makes the feedback obvious:
ranking → attention → behavior → logged data → training → ranking
The model is not simply estimating a fixed preference distribution. It participates in producing the future distribution it will later measure.
This is one of the deepest things I take from cybernetics:
Prediction stops being causally neutral once it is allowed to act on the world.
So the normative question is not a second essay stapled onto the first. It is what appears when the first argument is followed all the way through. If intelligence is realized in a loop, then the purposes around which that loop is organized—and the people whose states it regulates—belong inside the analysis too.
Where this leaves my view of NeuroAI
The worldview I am converging on is not that representation is unimportant, or that every intelligence problem should be rewritten as classical control theory.
It is that representation is only one slice through an intelligent process.
A brain state matters because of where it came from, what it makes possible, how it changes under perturbation, what action uses it, what consequence follows, and how that consequence changes the system next time.
That is why I want brain-likeness to mean more than similarity of internal geometry:
behavior + representation + dynamics + control + adaptation
For biological intelligence, brain + body + environment is not merely context around computation. It is part of the mechanism by which computation acquires meaning and consequence. For artificial intelligence, the same issue becomes harder to ignore as models gain tools, memory, persistent state, and the ability to intervene in the world.
This is the cybernetic version of the boundary claim in The System Is the Model. The relevant object may be larger than one parameterized function because models, memory, bodies, tools, other agents, and environments can occupy different causal positions inside the same feedback process.
The question I would like NeuroAI to ask more often is:
What computational organization allows an embodied agent to estimate, predict, act, learn, and remain adaptive in a changing world?
I mean the answer below as a research perspective, not as a necessary-and-sufficient definition of intelligence:
Intelligence is not merely a property of a model. It is a property of an organized system that uses information to predict, act, learn, and regulate itself through feedback.
The ability to optimize a purpose is not the wisdom to choose one.
The first changes where I look for intelligence.
The second is why finding it is not the end of the problem.
References and further reading
- Ashby, W. R. (1952). Design for a Brain. John Wiley.
- Ayzenberg, V., Bonner, M. F., & Bayet, L. (2026). Advancing NeuroAI through developmental alignment. Neuron.
- Cannon, W. B. (1932). The Wisdom of the Body. W. W. Norton.
- Cisek, P. (2007). Cortical mechanisms of action selection: The affordance competition hypothesis. Philosophical Transactions of the Royal Society B, 362, 1585–1599.
- Conant, R. C., & Ashby, W. R. (1970). Every good regulator of a system must be a model of that system. International Journal of Systems Science, 1, 89–97.
- D’Angelo, G., Pedersen, J. E., Hassan, T., et al. (2026). A benchmarking framework for embodied neuromorphic agents. Nature Machine Intelligence, 8, 300–312.
- Feather, J., Khosla, M., Murty, N. A. R., & Nayebi, A. (2025). Brain-Model Evaluations Need the NeuroAI Turing Test. arXiv.
- Friston, K. (2010). The free-energy principle: A unified brain theory?. Nature Reviews Neuroscience, 11, 127–138.
- Friston, K., FitzGerald, T., Rigoli, F., Schwartenbeck, P., O’Doherty, J., & Pezzulo, G. (2016). Active inference and learning. Neuroscience & Biobehavioral Reviews, 68, 862–879.
- Hassabis, D., Kumaran, D., Summerfield, C., & Botvinick, M. (2017). Neuroscience-inspired artificial intelligence. Neuron, 95, 245–258.
- Krakauer, J. W., Ghazanfar, A. A., Gomez-Marin, A., MacIver, M. A., & Poeppel, D. (2017). Neuroscience needs behavior: Correcting a reductionist bias. Neuron, 93, 480–490.
- Rao, R. P. N., & Ballard, D. H. (1999). Predictive coding in the visual cortex: A functional interpretation of some extra-classical receptive-field effects. Nature Neuroscience, 2, 79–87.
- Rosenblueth, A., Wiener, N., & Bigelow, J. (1943). Behavior, purpose and teleology. Philosophy of Science, 10, 18–24.
- Sadeh, S., & Clopath, C. (2025). The emergence of NeuroAI: Bridging neuroscience and artificial intelligence. Nature Reviews Neuroscience, 26, 583–584.
- Schrimpf, M., Kubilius, J., Hong, H., Majaj, N. J., Rajalingham, R., Issa, E. B., et al. (2018). Brain-Score: Which artificial neural network for object recognition is most brain-like?. bioRxiv.
- Todorov, E., & Jordan, M. I. (2002). Optimal feedback control as a theory of motor coordination. Nature Neuroscience, 5, 1226–1235.
- Wiener, N. (1948). Cybernetics: Or Control and Communication in the Animal and the Machine. MIT Press edition.
- Wiener, N. (1950/1954). The Human Use of Human Beings: Cybernetics and Society. Revised edition, 1954.
- Wolpert, D. M., Ghahramani, Z., & Jordan, M. I. (1995). An internal model for sensorimotor integration. Science, 269, 1880–1882.
- Zador, A. M., Escola, S., Richards, B., Ölveczky, B., Bengio, Y., Boahen, K., et al. (2023). Catalyzing next-generation artificial intelligence through NeuroAI. Nature Communications, 14, 1597.