Can Intelligence Be Learned from Its Own Footprints?
Beyond the Stochastic Parrot: Rethinking Artificial Intelligence
Cristiano Sacchi
Intelligence as a Function
On Human Thought, Artificial Intelligence, and the Space of Possible Answers
An exploratory essay
1. The Library of Babel
In his short story The Library of Babel, Jorge Luis Borges imagines a library containing every possible book that can be constructed from a fixed alphabet, using a fixed number of characters.
The library contains every book ever written, every book that will ever be written, and every book that could conceivably be written. It contains accurate histories of events that have not yet occurred, scientific discoveries that humanity has not yet made, and biographies of people who have never existed.
It also contains an almost inconceivable quantity of nonsense.
For every meaningful book, there are astronomical numbers of books containing meaningless sequences of characters. There are also books that differ from meaningful ones by a single letter, books containing plausible but incorrect information, and books whose contents are almost meaningful but ultimately incoherent.
The library is a literary illustration of combinatorial explosion.
Suppose we have an alphabet containing 30 possible characters and wish to produce a text of 10,000 characters. The number of possible texts is:
This number is so enormous that ordinary comparisons cease to be useful.
Yet human beings routinely produce meaningful sequences of thousands of characters.
When writing an essay, a letter, a scientific paper, or even a brief message, we select one sequence from this astronomical collection of possibilities.
We do so without explicitly considering more than an insignificant fraction of the alternatives.
This observation suggests an interesting starting point for thinking about intelligence.
Perhaps one manifestation of intelligence is the capacity to navigate an enormous space of possible outputs and produce something appropriate to a particular context and purpose.
Consider what happens when someone writes a sentence.
At every character, numerous continuations are possible. Most would lead to nonsense. Many others would produce grammatically correct but irrelevant statements. A smaller subset would preserve the intended meaning and direction of the argument.
The writer does not consciously enumerate these possibilities. Somehow, knowledge, experience, intentions, linguistic competence, and other mental processes guide the production of a coherent sequence.
The result is remarkable when viewed against the combinatorial background.
A meaningful paragraph is an extraordinarily constrained selection from a vast possibility space.
Figure 1 — A meaningful text is one selection from an astronomical space of possibilities.
This need not be a complete definition of intelligence. Indeed, intelligence manifests itself in many activities that have nothing to do with language. But the example provides a relatively simple environment in which to examine a more general phenomenon.
It also leads to a question: could intelligence be understood, at least in part, as a function that maps circumstances into appropriate outcomes?
2. Intelligence Beyond Language
Writing provides a convenient example because its outputs are discrete, observable, and easy to represent mathematically.
But intelligent behavior is hardly limited to writing.
Consider driving a car.
A driver continuously receives information about the road, other vehicles, speed, direction, visibility, and numerous other circumstances.
In response, the driver produces a sequence of actions: steering, accelerating, braking, changing lanes, and adjusting attention.
The range of possible actions is enormous. Only a relatively restricted portion is appropriate at any given moment.
An experienced driver negotiates this complexity without necessarily conducting explicit, conscious reasoning.
Indeed, much of the activity becomes automatic with experience.
The driver's behavior is nevertheless intelligent in an ordinary and meaningful sense. It requires perception, learning, judgment, adaptation, and the ability to produce appropriate actions under changing circumstances.
We could make similar observations about playing a musical instrument, designing a building, solving a mathematical problem, conducting a negotiation, or improvising a conversation.
The inputs and outputs differ, but a common structure appears:
A system receives information about its circumstances and produces an outcome appropriate to those circumstances.
In the broadest mathematical sense, we can represent this as a function:
where is a set of possible inputs and is a set of possible outputs.
An input might include sensory information, accumulated knowledge, goals, and context. An output might be a string of text, a physical action, a design, a decision, or some combination of these.
Of course, describing intelligence as a function is not, by itself, particularly informative. Almost any deterministic system can be represented as a function.
A conventional thermostat maps measured temperature to an action. A calculator maps numerical inputs to numerical outputs.
Neither is generally considered intelligent in the same sense as a human being.
The interesting question is not merely whether a mapping exists. It is what distinguishes mappings capable of producing intelligent behavior from those that are not.
Before addressing that question, however, it is worth examining a fundamental distinction between a mathematical function and the process used to compute it.
3. A Function Is Not Its Implementation
Two computational systems can perform entirely different operations while implementing the same mathematical function.
This is a straightforward mathematical observation, but its implications for discussions of artificial intelligence are considerable.
Suppose two functions, and , have the same domain and codomain, and satisfy:
for every possible input in their shared domain.
They are, mathematically, the same function.
The procedures used to evaluate them may be entirely different.
One algorithm might compute a result through a sequence of algebraic operations. Another might use a numerical approximation. A third might retrieve a previously computed result from a table.
If their mappings are identical, the functions are identical, even if their implementations are not.
Fourier analysis provides a familiar illustration of the distinction between a function and its representation.
Under appropriate mathematical conditions, a signal can be represented through a combination of frequency components. The signal and its Fourier representation describe the same mathematical object through different structures.
Similarly, numerical mathematics frequently allows us to compute useful approximations to functions whose solutions cannot conveniently be expressed in closed analytical form.
We can evaluate such functions without possessing a compact symbolic expression that makes their behavior immediately understandable.
This distinction between understanding a function, representing it, and evaluating it is important.
We should not assume that reproducing a function requires reproducing the original mechanism through which it was first realized.
And we should not assume that a computational implementation is incapable of exhibiting a particular functional property simply because its elementary operations appear unrelated to that property.
This brings us to a thought experiment inspired by Borges.
4. A Library Containing Every Answer
Imagine constructing an enormous library containing the answers that a particular person would give to every possible question of up to 100 characters.
Let us assume that the person, their knowledge, and their relevant mental state are fixed at a particular moment.
For each possible question, we ask the person for an answer and record it.
The number of possible questions is astronomical, and the construction is clearly impossible in practice. But impossibility of physical construction does not prevent us from considering the mathematical consequences.
Eventually, our hypothetical library contains a complete mapping from every permitted question to the answer that person would produce.
Now imagine replacing the person with a computer that simply looks up the appropriate answer.
The computer performs no reasoning in the conventional sense.
It does not analyze the question, construct an argument, evaluate alternatives, or draw conclusions.
It simply retrieves a stored string.
Nevertheless, for every question within the specified domain, it produces exactly the answer that the person would have produced.
Let be the person's answer to question , and be the corresponding library lookup.
By construction:
where is the complete set of permitted questions.
Within this domain, the two systems implement the same function.
The human brain and the lookup table have radically different implementations, but their observable input-output behavior is identical.
Figure 2 — Different computational mechanisms can implement an identical input-output mapping.
Would we call the library intelligent?
The answer depends on what we mean by intelligence.
If intelligence requires a particular internal process, the library may fail to qualify.
But if intelligence is understood functionally, through the capacity to produce intelligent responses, denying the library that property becomes more difficult.
After all, it produces precisely the responses of the intelligent person it represents.
The point becomes clearer if we distinguish intelligent behavior from intelligent thinking.
The lookup table provides a particularly stark example of a system that can exhibit the former without performing anything recognizable as the latter during retrieval.
It is possible to object that the intelligence resides in the human who originally produced the answers, rather than in the library itself.
That is a reasonable philosophical position. But it also reveals an additional assumption: intelligence must involve more than functional behavior.
That assumption may ultimately be justified, but it does not follow from the mathematics of input-output mappings.
Our thought experiment does not settle the philosophical question. It merely makes the disagreement explicit.
There are limitations to the example. The table is fixed, does not learn, and cannot respond outside its specified domain. Moreover, if responses depend on previous exchanges, emotional states, or changing circumstances, these must also be represented in the input.
The hypothetical could be extended to include such information, but the resulting library would become even more impossibly large.
For our purposes, the essential observation remains: the same intelligent input-output behavior can, in principle, be implemented through mechanisms that share almost nothing in common.
This thought experiment bears a close resemblance to John Searle's celebrated Chinese Room argument, introduced in 1980. Searle imagined a person who does not understand Chinese but follows detailed instructions for manipulating Chinese symbols, producing responses that appear meaningful to a Chinese-speaking observer. His argument was that successfully manipulating symbols according to formal rules does not, by itself, establish genuine understanding.
Our hypothetical lookup library raises a related question from a different direction. Rather than attempting to establish whether the system understands its answers, we ask whether identical input-output behavior constitutes functional equivalence, regardless of how that behavior is produced. The lookup library does not refute Searle's argument, nor does Searle's argument invalidate the mathematical equivalence of the mappings. The disagreement concerns whether functional equivalence is sufficient to attribute intelligence or understanding to the system. This is precisely the philosophical distinction our thought experiment is intended to expose.
This is consistent with the broad philosophical tradition of functionalism, although it does not require accepting every claim associated with that tradition.
5. A Preliminary Functional View of Intelligence
We can now return to the distinction between an arbitrary mapping and an intelligent mapping.
The Library of Babel contains all possible texts of a specified length. Yet only a restricted subset constitutes appropriate answers to a particular question.
Suppose we call this restricted collection the intelligent output subset.
For an input , let represent the set of outputs that would qualify as intelligent responses to that input.
Then:
An intelligence function would aim to produce an output satisfying:
This formulation is deliberately modest.
It does not tell us how to determine whether an answer belongs to . Nor does it explain how the function selects an answer.
It merely proposes a mathematical structure in which those questions can be discussed.
There is an important consequence: an intelligent answer need not be unique.
Two people given the same problem may produce different solutions, both of which are intelligent.
Two architects may design very different buildings that satisfy the same requirements. Two mathematicians may establish the same theorem through different proofs. Two writers may offer distinct and equally illuminating interpretations of a novel.
Intelligence does not necessarily mean finding one predetermined answer.
It may mean finding one satisfactory outcome among many.
Nor does an intelligent system necessarily need to identify the optimal outcome.
In many practical situations, finding a solution that satisfies the relevant constraints is sufficient.
Human reasoning frequently proceeds heuristically, using experience, approximation, and learned patterns rather than exhaustive examination of every possibility.
But heuristic computation should not be made a necessary condition of intelligence.
Our lookup-table example demonstrates why.
The library implements an entirely algorithmic retrieval procedure while reproducing responses originally produced through human intelligence.
If intelligence is a functional property, its implementation need not itself use heuristics.
We might therefore propose the following preliminary definition:
Intelligence is the capacity of a system to transform information about its circumstances into contextually appropriate, goal-directed outcomes across a sufficiently broad and varied range of situations.
This is a working definition rather than a definitive one.
It leaves several important questions unresolved, particularly the meaning of appropriate and the breadth of circumstances required.
The intelligent subset is also dependent on the input. A response that is intelligent in one context may be inappropriate in another.
Consider humor.
The same joke may be amusing at a dinner party and deeply inappropriate at a funeral.
The words are identical, but their suitability to the situation differs. The intended goal, social expectations, and surrounding circumstances affect whether the response qualifies as intelligent.
One possible direction for developing the definition would involve examining how successfully an outcome satisfies the goals associated with its input.
But goal satisfaction alone is not enough. A random action may occasionally achieve a goal through luck.
For now, it seems preferable to leave the precise characterization of the intelligent output subset as an open question.
The proposed framework can still be useful without fully resolving it.
6. Large Language Models and the Next-Token Objection
This functional perspective provides an interesting way to examine large language models.
A common description of an LLM is that it predicts and generates one token at a time.
For conventional autoregressive language models, this is technically correct.
The model receives a sequence of tokens, computes a probability distribution over possible continuations, selects another token, and repeats.
Its output is constructed sequentially.
This mechanism has sometimes contributed to the characterization of LLMs as stochastic parrots: systems that reproduce statistical patterns in language without possessing intelligence or understanding.
The original stochastic-parrot critique is broader than the mechanics of token prediction. It raises questions about meaning, grounding, training data, and the social implications of language technologies.
Those questions deserve consideration on their own terms.
But one narrower argument merits examination: the suggestion that because an LLM generates individual tokens through statistical computation, its overall behavior cannot be intelligent.
That conclusion does not follow from the description of its mechanism.
When a person writes an essay, the text also emerges sequentially.
A writer types one character after another, or speaks one word after another. The person may have a general plan, but often does not know the exact wording of a sentence several lines ahead.
We do not ordinarily evaluate the intelligence of the writer by examining individual keystrokes.
We evaluate the resulting argument, explanation, or creative work.
A similar distinction applies to numerical computation.
An iterative algorithm may require millions of elementary operations to approximate the solution to a mathematical problem.
We do not describe the algorithm's mathematical purpose solely in terms of its individual iterations.
We describe it by the function it computes and the result it produces.
For an LLM, we can similarly distinguish the token-generation mechanism from the complete input-output mapping.
Let:
where:
- represents the prompt and available context;
- represents the trained parameters;
- represents the relevant generation settings and, where applicable, random choices;
- is the complete generated response.
The trained model implements a mapping from input conditions to complete outputs.
When randomness is involved, we can equivalently describe the model as defining a probability distribution over complete responses.
Token-by-token generation is how the model evaluates that mapping.
It is not necessarily the most useful level at which to assess the intelligence of its behavior.
This does not mean token-by-token generation is irrelevant. The sequential architecture influences what models can do, how they plan, and where they encounter limitations.
It does mean that a description of the elementary computational procedure does not, by itself, establish the intellectual capabilities of the complete system.
The meaningful question is what the resulting function can accomplish.
7. Training as an Inverse Problem
How does an LLM acquire such a function?
One useful way to think about training is through the mathematical concept of an inverse problem.
In a conventional forward problem, we know a model and use it to calculate its outputs.
In an inverse problem, we observe outputs and attempt to infer a model capable of producing them.
Much of scientific modeling involves this distinction.
We observe the behavior of a system, collect measurements, and construct mathematical representations that reproduce those observations.
The resulting model need not share the physical mechanism of the system being studied.
And it need not provide an analytically simple description of that mechanism.
Something broadly similar happens during LLM training.
Human civilization has produced an extraordinary body of intellectual work: literature, mathematics, scientific explanations, philosophical arguments, technical documents, conversations, computer programs, and countless other products of human activity.
These are not merely collections of facts.
They are also observable products of intelligent behavior.
They contain examples of reasoning, abstraction, explanation, argumentation, analogy, creativity, and problem-solving.
A language model is trained to predict linguistic continuations from this material. Through that process, its parameters are adjusted to represent relationships useful for making those predictions.
Although the training objective is not explicitly to reconstruct human intelligence, successfully modeling complex human-produced material can require acquiring capabilities that resemble aspects of the processes through which that material was produced.
In this sense, we can interpret training as constructing a function that approximates certain observable manifestations of human intelligence.
This is an interpretation of what training accomplishes, not a claim that its formal objective is to recover a uniquely defined human intelligence function.
The distinction matters because inverse problems are frequently underdetermined.
Different underlying functions may explain the same observed data.
Suppose we are given a collection of points sampled from an unknown mathematical function.
We might construct a polynomial, a spline, or a neural network that reproduces those observations.
All may fit the sample well while behaving differently elsewhere.
The observations do not necessarily identify one unique underlying function.
The same issue arises with intelligence.
There is no reason to assume that one particular computational process is uniquely determined by the historical products of human thought.
An LLM may learn a function that reproduces many intelligent behaviors while differing substantially from the biological processes that originally generated its training material.
This is not a defect in the analogy. It is central to it.
The aim of functional reconstruction need not be to reproduce the original mechanism. It may be sufficient to construct a different mechanism capable of generating comparable intelligent outcomes.
Whether two such systems are fully functionally equivalent is an empirical question, not something established merely by their ability to produce impressive answers.
8. From Samples to New Intelligent Outcomes
The distinction between memorizing examples and learning a function is essential.
Return to our mathematical example.
Suppose we construct a spline through a collection of sampled points.
Once the spline is defined, we can evaluate it at inputs that were never included in the original sample.
The interpolated function can generate new values.
Figure 3 — A learned function can generate values beyond its observed samples.
These values need not be stored anywhere in advance.
This provides an analogy for what happens when a trained model generates novel answers.
The historical body of human knowledge contains an immense collection of products of intelligent activity.
But those products do not exhaust the possibilities of intelligent behavior.
There are mathematical proofs not yet discovered, engineering designs not yet imagined, scientific explanations not yet formulated, and literary works not yet written.
If we regard the historical record as a sample of the intelligent output space, then it is a large but incomplete sample.
A learned intelligence function may generate outputs beyond those observations.
Indeed, the ability to do so is one of the most interesting features of modern AI.
A model may combine concepts in ways not explicitly represented in any particular training example, derive new conclusions, or propose solutions to previously unencountered problems.
Not every novel output is correct, useful, or genuinely original. Novelty alone is not evidence of intelligence.
Nevertheless, the capacity to generate appropriate outcomes that were not explicitly present in the training material distinguishes learning a generalizable function from merely retrieving stored answers.
The spline analogy has limitations.
Language and reasoning do not occupy a simple, continuous space in which interpolation has an obvious geometric meaning. Furthermore, extrapolation outside well-represented regions can be unreliable.
But the central point survives:
A function inferred from samples can possess capabilities that are not reducible to the reproduction of those samples.
This also helps clarify the distinction between knowledge and intelligence.
Knowledge includes information about the world and previously established solutions.
Intelligence includes the capacity to use information to generate appropriate responses, potentially including responses never previously observed.
The two are closely related, but conceptually distinguishable.
9. Knowledge, Intelligence, and Education
Human experience offers a useful illustration of this distinction.
Consider two individuals.
One possesses substantial general reasoning ability but relatively little accumulated knowledge.
The other has more modest reasoning ability but extensive education, excellent memory, and considerable experience.
In many circumstances, the second person may appear more intelligent.
They can draw on facts, established methods, examples, and accumulated understanding that the first person does not possess.
Yet when confronted with a genuinely unfamiliar problem for which all relevant information is supplied, the first person might reason more effectively.
This suggests that knowledge and reasoning capability contribute differently to observable intellectual performance.
An informal formulation would be:
Knowledge helps determine what an intelligent system can answer. Intelligence helps determine what it can do with what it knows.
The distinction should not be pushed too far, however.
Knowledge acquisition can also change reasoning capability itself.
Imagine a person who has spent years studying literature but has little mathematical education.
They begin studying mathematics, eventually encountering hundreds or thousands of proofs.
At first, they learn definitions, theorems, and established results.
But something else may happen.
Repeated exposure to mathematical reasoning can change how they approach problems.
They learn to recognize structural similarities, formulate abstractions, construct proofs, and distinguish valid arguments from invalid ones.
They become capable of producing mathematical solutions that they have never previously encountered.
In the language of our functional framework, education has not merely supplied additional information.
It has modified the mapping through which the person transforms problems into solutions.
The person has acquired new capabilities through exposure to examples from a previously unfamiliar class of intelligent behavior.
An analogous process plausibly occurs during LLM training.
Exposure to mathematical texts may teach a model more than mathematical facts. It may also provide patterns of mathematical reasoning.
Exposure to engineering literature may provide approaches to evaluating constraints and trade-offs.
Exposure to legal arguments may provide patterns of interpretation and argumentation.
The extent to which a model acquires these capabilities, rather than merely reproducing familiar examples, remains an empirical question.
But the analogy with human education is instructive.
Education shapes intellectual dispositions
There is a further observation worth considering.
Different educational and professional histories tend to produce recognizable differences in how people approach problems.
Mathematicians often seek abstractions, formal relationships, and rigorous demonstrations.
Engineers often emphasize constraints, trade-offs, feasibility, and workable solutions.
Writers may pay particular attention to language, interpretation, human motivation, and narrative structure.
These are tendencies, not rigid boundaries.
An engineer can reason mathematically. A mathematician can approach a problem pragmatically. A writer can offer an excellent engineering insight.
Nevertheless, sustained exposure to a discipline can shape a person's intellectual habits.
In our functional picture, education influences which regions of the space of possible intelligent responses an individual is more inclined to explore.
Two people may receive the same input and produce different intelligent outputs, partly because their educational histories have shaped their respective mappings.
Their functions differ not only in the knowledge they contain, but also in the dispositions they express.
This observation has a potentially interesting implication for LLMs.
A single model trained on material from many intellectual traditions may acquire overlapping patterns of reasoning from those traditions.
It may be capable of approaching the same problem mathematically, technically, philosophically, or creatively.
It may also combine approaches in ways that would be less natural for an individual whose education has been concentrated in one field.
This does not guarantee that the model has mastered every discipline represented in its training material.
But it suggests that the diversity of training observations may matter in a particularly significant way.
Training can shape not only the quantity of information available to a model, but also its characteristic ways of producing solutions.
10. Are There Different Classes of Intelligence?
The preceding discussion suggests a more speculative possibility.
Perhaps different classes of intelligent behavior can be learned from different kinds of observations.
The distinction need not correspond neatly to conventional academic disciplines.
Mathematics, literature, engineering, and philosophy may share substantial underlying structures.
The more interesting divisions may arise where the relationships between inputs, goals, and successful outputs are fundamentally different.
Consider a model trained on the entire history of human musical composition.
Suppose it becomes extraordinarily capable of generating new music.
It might produce compositions unlike anything written before, combining musical structures in original ways.
Now imagine asking it to compose music for dogs, using ultrasonic frequencies outside normal human hearing.
The problem is not simply to produce human music at a higher pitch.
We would first need to determine whether dogs experience anything analogous to musical appreciation, what acoustic patterns matter to them, and what outcomes would constitute success.
Human musical history may contain little direct evidence about those relationships.
A model that has reconstructed many principles of human musical composition would not necessarily have reconstructed the function required to produce meaningful ultrasonic music for dogs.
It might still succeed through additional scientific knowledge, experimentation, simulation, or feedback.
But the original body of human musical compositions would not, by itself, guarantee that capability.
This illustrates a possible limitation of learning intelligent behavior from a particular class of observations.
The issue is not that the system cannot generate novel outcomes.
It may generate entirely new outcomes within the class of behavior it has learned.
The question is whether observations from one class contain enough information to infer the relationships required for a substantially different class.
The mathematical concept of identifiability is relevant here.
If multiple candidate functions are equally consistent with the observed data but behave differently in an unobserved domain, those observations alone cannot uniquely determine which candidate will behave appropriately there.
Additional assumptions or evidence are needed.
This suggests the following hypothesis:
The classes of intelligent behavior that can reliably be learned from a body of observations may be constrained by the kinds of relationships represented in those observations, rather than simply by their quantity.
This is not a derived theorem about LLMs.
It is a conjecture suggested by the functional interpretation of training.
It also should not be confused with the claim that an LLM can only reproduce things already found in its training data.
The two propositions are quite different.
A function can generate entirely novel results while still being limited in the classes of relationships it has learned to represent.
That distinction may be useful when considering the future development of artificial intelligence.
Current language models already operate across an extraordinarily broad range of intellectual domains. Multimodal models extend their observations beyond text, and systems incorporating tools, experimentation, and environmental feedback can acquire other kinds of evidence.
The functional perspective suggests one possible reason why these developments matter.
They may provide information about classes of intelligent behavior that cannot be reliably inferred from textual observations alone.
The limitation, in this interpretation, is not necessarily that language models use the wrong computational mechanism.
It is that some intelligent capabilities may require kinds of experience or evidence not adequately represented in the observations from which the function was learned.
11. More Data, More Knowledge, or More Intelligence?
A second speculative consequence concerns the effects of increasing training data.
Modern LLMs are trained on enormous quantities of information, and increasing the scale of training has produced substantial improvements in their capabilities.
But from the perspective developed here, we can ask whether every additional training example contributes equally to the reconstruction of intelligent behavior.
Consider again the problem of approximating an unknown function from samples.
Early observations may reveal important features of the function that were previously unknown.
As more observations accumulate, the approximation improves.
Eventually, however, additional samples from already well-represented regions may provide relatively little new information about the underlying structure.
The approximation may continue improving, but with diminishing returns.
Could something analogous happen with LLM training?
Perhaps.
We can imagine a point at which a model has learned much of the generalizable reasoning structure available from a particular class of training material.
Additional data might continue expanding its knowledge without producing proportionally large improvements in its general reasoning capabilities.
A model might become much more knowledgeable about medicine, history, engineering, or literature while its ability to perform abstract reasoning improves more slowly.
It would become more capable in many practical situations, but some of that improvement would arise from having more information available rather than from a corresponding improvement in general reasoning.
This is not a claim that knowledge and reasoning are physically separated within a neural network.
They may be deeply intertwined in the learned parameters.
It is a conceptual distinction between two possible effects of training.
There is also an important qualification.
Additional data may reveal new reasoning structures, not merely new facts.
As the example of mathematical education suggests, learning a genuinely unfamiliar domain can change the way a system approaches problems.
Therefore, diminishing returns within one well-represented class of behavior do not necessarily imply diminishing returns from exposure to new classes.
This raises a possibility worth investigating:
The quality, diversity, and informational novelty of training observations may eventually matter more for some capabilities than simply increasing the total quantity of data.
A relatively small number of examples revealing new reasoning patterns might provide greater improvement in certain capabilities than a much larger collection of examples reinforcing familiar ones.
This possibility is consistent with the importance of data quality already recognized in machine learning, but the proposed distinction between functional refinement and knowledge acquisition remains hypothetical.
There are many other factors: model architecture, representational capacity, optimization methods, inference-time computation, and the interaction between data and learning procedures.
No universal threshold of diminishing returns follows from our framework.
Still, the hypothesis provides a way to think about a practical question: whether future advances in AI will depend primarily on collecting more examples of existing human intellectual activity or on discovering better ways to learn from more varied kinds of evidence.
12. The Problem of Evaluating Intelligent Outcomes
Our framework deliberately leaves one question unresolved.
What makes a particular output intelligent?
We have proposed an intelligent subset of possible outputs, but we have not specified a mathematical procedure for identifying its members.
This is not a minor technical detail.
It is arguably one of the most difficult questions in any attempt to formalize intelligence.
A correct mathematical proof may be relatively straightforward to evaluate once its premises and formal rules are established.
An engineering solution can be assessed against performance requirements.
But what makes a philosophical argument insightful, a joke funny, a poem moving, or a strategic decision wise?
Such evaluations depend on circumstances, goals, available knowledge, and often human judgment.
Furthermore, intelligence and correctness are not identical.
An intelligent person can reason carefully and arrive at an incorrect conclusion.
A simple lookup table can produce a correct answer without performing any reasoning during retrieval.
Historically, scientists working within Newtonian physics could reason intelligently while using a physical framework later shown to be incomplete.
Newtonian mechanics remains extraordinarily effective within its applicable domain. Its limitations do not retroactively make the intellectual activity of earlier physicists unintelligent.
Similarly, an AI system may produce an incorrect output for several different reasons.
It may have learned incorrect information. It may lack relevant evidence. It may fail to generalize appropriately, or it may make an error in constructing an inference.
The broad label hallucination is often used for AI outputs that present unsupported or false information with apparent confidence. The label describes a class of observable failures without necessarily identifying their causes.
For the present argument, the important point is that a system's intelligence should not be equated with the correctness of every individual output.
Nor should isolated correct answers establish intelligence.
What matters is a system's capacity to produce appropriate outcomes across a sufficiently broad and varied set of circumstances.
How that capacity should be evaluated remains open.
We might eventually consider goal satisfaction, reliability, generalization, creativity, or other criteria.
We might also conclude that intelligence cannot usefully be represented by a single numerical measure.
But it is not necessary to solve those problems in order to examine the implications of treating intelligence as a functional capability.
Mathematics often allows us to study the relationships between objects whose complete characterization remains unfinished.
We can introduce the intelligent output subset as a working concept while acknowledging that its boundaries are not yet rigorously defined.
13. A Different Perspective on Artificial Intelligence
We can now return to the question that motivated this exploration.
What should we make of the claim that LLMs are merely statistical systems generating successive tokens?
The description is correct at one level of implementation, but it does not adequately describe the complete function that the system performs.
A language model receives contextual information and produces responses that can exhibit coherence, abstraction, reasoning, creativity, and adaptation across an unusually broad range of problems.
Its computational mechanism differs substantially from the operation of biological brains.
We do not need to deny those differences.
Indeed, one of the principal advantages of a functional perspective is that it allows us to acknowledge them without treating them as automatic evidence against intelligence.
The same mathematical function can, in principle, be computed by different mechanisms.
Conversely, two systems may implement different functions while exhibiting overlapping intelligent capabilities.
There is no need to claim that artificial intelligence has reconstructed the human brain.
Nor is there a need to claim that it has reconstructed one uniquely defined function representing all of human intelligence.
A more modest interpretation is sufficient.
LLM training constructs a computational function from a vast collection of observations, many of which are products of human intellectual activity.
That function can generalize beyond the observed examples and produce new outputs that, in numerous circumstances, satisfy the criteria we ordinarily associate with intelligent behavior.
The learned function is not necessarily identical to any human intelligence function.
It need not be.
The more relevant question is whether its outputs demonstrate the relevant capabilities across the contexts in which it operates.
This also clarifies why the Library of Babel thought experiment matters.
Our hypothetical lookup table could reproduce the intelligent responses of a human while performing no reasoning during retrieval.
A trained LLM is very different from that table. It learns a compact, generalizable computational structure rather than explicitly storing every possible question-answer pair.
But the thought experiment establishes the conceptual distinction between intelligent behavior and the particular mechanism that produces it.
Once that distinction is accepted, the claim that a system cannot be intelligent because its implementation consists of statistical or numerical operations requires an additional argument.
The mechanism alone does not settle the question.
One might still argue that genuine intelligence requires consciousness, subjective understanding, intentionality, embodiment, or some other property.
These are serious philosophical possibilities.
But they are additional requirements, not consequences of the fact that a system predicts tokens.
Our functional framework does not attempt to resolve consciousness or subjective experience.
It addresses the narrower question of whether intelligent behavior can be understood as a computationally implementable mapping from circumstances to appropriate outcomes.
14. Some Concluding Thoughts
We began with Borges's imaginary library.
It contains an astronomical number of possible texts, almost all of them useless for any particular purpose.
Human intelligence allows us to produce meaningful work without exhaustively exploring those possibilities.
This observation suggested a functional interpretation: intelligence may be understood as the capacity to map inputs into appropriate outcomes selected from a vast space of possibilities.
From there, several ideas followed.
A mathematical function is distinct from the procedure used to evaluate it.
Different implementations can exhibit identical input-output behavior.
Training a language model can be interpreted as an inverse problem: constructing a function from observations of human-produced material.
The historical products of human intelligence provide a rich but incomplete sample of intelligent outcomes.
Learning from those observations may allow a system to generate entirely new intelligent outputs, rather than merely reproduce previously encountered answers.
Knowledge acquisition can improve performance, but exposure to examples of intelligent behavior may also shape the underlying function, developing new capabilities and intellectual dispositions.
And the kinds of observations available during training may influence the classes of intelligent behavior that can reliably be learned.
These ideas do not amount to a complete mathematical theory of intelligence.
In particular, we have not established a rigorous criterion for membership in the intelligent output subset. We have not demonstrated that all intelligent behavior can be characterized through one unified function, or that artificial and biological intelligence implement the same mapping.
Nor have we established the speculative propositions concerning the limits of learning from particular classes of data or the eventual marginal value of additional training examples.
Those remain hypotheses.
The aim has been more limited: to explore whether a functional perspective provides a useful way to think about intelligence, and particularly about the relationship between human intelligence and large language models.
One consequence of that perspective seems worth emphasizing.
Human intelligence need not be regarded as the defining implementation of intelligence.
It may instead be one particular realization of a more general capability.
Artificial intelligence may constitute another realization, with overlapping but different strengths, limitations, and modes of operation.
The important distinction is between what a system is capable of doing and how it accomplishes it.
This does not mean that mechanisms are unimportant. They matter enormously for reliability, efficiency, learning, adaptation, transparency, and many other practical considerations.
But the mechanism through which a function is computed should not be confused with the functional behavior it produces.
Perhaps the most useful question, then, is not whether artificial intelligence thinks in the same way humans think.
It almost certainly does not, at least in many important respects.
The question is whether different computational and biological systems can exhibit a common functional capability that we recognize as intelligence.
If they can, then intelligence may be better understood not as the exclusive property of a particular biological mechanism, but as a property that different systems can realize through different means.
And perhaps the enormous space of intelligent possibilities contains regions that neither human civilization nor contemporary artificial intelligence has yet explored.
The Library of Babel, in Borges's imagination, contains all possible books.
Our real intellectual history contains only those we have managed to write.
The distinction between those two collections may be a useful way to think about both the achievements and the possibilities of intelligence.
Intelligence is not the library.
It is what allows us to find, construct, and perhaps discover meaningful things within an otherwise overwhelming universe of possibilities.