We Used to Find the Pattern First
How artificial intelligence changes prediction, discovery, and what humans can know
The pattern is not a line written by a programmer. It is a terrain shaped by error.
The historical break
Earlier information technology required the human to discover the pattern, understand it, and convert it into instructions.
Artificial intelligence can discover and exploit predictive structure that no human first had to understand or formalize.
We used to give the machine our understanding. Now we can give it experience and ask it to form a useful model.
The complete edition / Two ways in
Read it. Hear it.
Preview the expanded textbook here before downloading it, or listen to the narrated edition without leaving the page.
Listen now
Press play to hear the narrated edition of We Used to Find the Pattern First.
Part I / The primitive
Beneath creation
The artifact is what the human sees. Prediction is what the machine does. To understand artificial intelligence, begin under the output.
CH / 01
The prediction underneath the artifact
Artificial intelligence does not begin by creating a book, a plan, an image, a diagnosis, a strategy, or a line of software. It begins by estimating what should come next.
We name the output by its visible form. A prediction that continues for sixty thousand words becomes a book. A prediction arranged around a business problem becomes a report. A prediction expressed as pixels becomes an image. A prediction written in Python becomes software. A prediction ordered into future actions becomes a plan.
The forms are different. The primitive is the same.
A language model estimates the probability of possible next tokens given the tokens already present. An image model estimates visual structure under a conditioning signal. A classifier estimates which label is most probable given an input. A forecasting system estimates a future value given prior observations. Even when the system produces a long chain, each local move is conditioned by the state created by the moves before it.
This does not make the output trivial. A cathedral is made one stone at a time. Its local construction does not erase its global form. In the same way, next-step prediction can accumulate into an argument, a program, an image, or an investigation. The important correction is not that the artifact is unreal. The correction is that creation is the external appearance of predictive computation.
That sentence places the burden where it belongs. The machine cannot predict from nothing. It needs structure that connects what has happened to what is likely to happen. The book, plan, image, or solution is the visible release of that structure.
Ask what an AI creates and you will receive a catalog of artifacts. Ask what operation makes those artifacts possible and the catalog collapses into one word: prediction.
The next question is therefore unavoidable.
What makes prediction possible?
CH / 02
What a pattern is
A pattern is structure that reduces uncertainty.
Suppose a light flashes red, blue, red, blue, red, blue. Before seeing the sequence, the next color could have been either. After observing the alternation, the next color is no longer equally uncertain. The history contains structure. The structure changes the probability of the next event.
That is enough for a working definition. A pattern need not be decorative. It need not repeat visibly. It need not fit into a sentence. It need not be simple enough for a human to notice. It must only make some outcomes more expected than others.
If every possible future remains exactly as likely after observing the past, the past supplied no predictive pattern. If the probabilities move, structure was present.
P(Y | X) ≠ P(Y)
Observing X changes the probability of Y. X contains predictive information about Y.Patterns come in many forms. A repeated sequence is a pattern. A statistical association is a pattern. An invariance under rotation is a pattern. A grammar is a pattern. A causal mechanism creates patterns, but causation is not required for prediction. A proxy can predict an outcome without producing it.
This last distinction matters. Prediction asks whether information reduces uncertainty about an outcome. Explanation asks why the outcome occurs. A model can succeed at the first while remaining weak at the second.
The observer is part of the claim
Pattern is never merely a property of data in isolation. It is a relation among data, representation, objective, and observer. The alternating lights are easy to represent because time gives them a natural order. A relationship spread across five hundred clinical variables may be just as real and far harder for a human to detect.
The phrase “there is no pattern” is therefore stronger than it sounds. It may mean there is no structure. It may mean we chose the wrong variables. It may mean we represented the variables badly. It may mean the structure is too weak for the available sample. It may mean the process changed. Or it may mean the observer lacks the capacity to find what is present.
No pattern, no prediction. But failure to predict does not prove the absence of pattern. It may reveal the boundary of the predictor.
CH / 03
Intelligence and compression
If a pattern exists, the data can be described more economically than by listing every observation independently.
Consider a thousand-character string made by repeating 10 five hundred times. The literal string is long. Its rule is short: “repeat 10 five hundred times.” The shorter description is possible because the sequence contains regularity.
Now consider a thousand truly random bits. There may be no description meaningfully shorter than the bits themselves. The sequence gives us nothing to factor out. In the language of algorithmic information theory, an incompressible string is random relative to the descriptive machinery being used.
This is why the claim “intelligence is compression” has real depth and also needs discipline. Compression is not merely shrinking a file. It is finding reusable structure. A successful compressor says: these observations are not isolated; this shorter model can regenerate or predict them.
Kolmogorov complexity expresses the idealized version. For a string x, its complexity is the length of the shortest program that outputs x on a fixed universal computer.
K(x) = minp: U(p)=x |p|
The shortest program p that makes universal machine U produce x. The exact value is not computable in general, but the concept defines the target.A short program reveals strong regularity. A long irreducible program reveals little exploitable structure. The fixed choice of universal machine matters by a constant, but not enough to destroy the central idea for sufficiently long descriptions.
Yet compression alone is not a complete definition of intelligence. A specialized compressor can be excellent on one distribution and useless elsewhere. Intelligence must also involve transfer: the discovered structure must support accurate prediction or successful action across situations that were not copied from the past.
A disciplined working definition is stronger:
Compression is evidence that structure was found. Prediction tests whether that structure reaches beyond the observations already seen. Action tests whether the prediction matters in an environment.
CH / 04
Compression revealed through time
Prediction and compression are not neighboring metaphors. For probabilistic sequences, they are operationally convertible.
If a model assigns high probability to the event that actually occurs, that event requires few bits to encode under an ideal code. If the model assigns low probability to the event, the event is surprising and requires more bits.
I(x) = −log2 p(x)
An event with probability 1/2 costs one bit. Probability 1/8 costs three bits. Accurate expectations shorten the message.For a sequence, the code length accumulates one conditional prediction at a time:
L(x1:n) ≈ Σt=1n −log2 p(xt | x<t)
The better the next-step probabilities, the shorter the total description.Arithmetic coding turns those probabilities into a compressed bitstream. Run the relationship in the other direction and a compressor becomes evidence of a predictive model. The 2023 paper “Language Modeling Is Compression” demonstrates the equivalence across text and other data types: powerful predictive models can act as general-purpose compressors, and compressors can support conditional generation.
This gives the book its central mathematical hinge:
The compressor says: “I have found enough structure that I do not need to transmit every observation independently.” The predictor says: “I have found enough structure that I can distribute probability over what comes next.” Both are reports about uncertainty reduced by a model.
Claude Shannon’s information theory supplied the quantitative language. Ray Solomonoff pushed the idea toward universal induction: prefer computable explanations according to their descriptive length, update them against observations, and let the weighted explanations predict the continuation. The ideal is mathematically profound and generally incomputable. Its importance is not that modern neural networks secretly implement Solomonoff induction. They do not. Its importance is that simplicity, probability, compression, and prediction can be joined inside one formal frame.
Once that connection is visible, a model no longer looks like a bag of answers. It looks like compressed predictive structure.
Part II / The reversal
The human had to know
Traditional software inherited the result of human intelligence. Machine learning changes where useful structure can originate.
CH / 05
The old order
For most of the history of information technology, intelligence stood upstream of the machine.
A human encountered a recurring situation. The human noticed a relationship. The human understood it well enough to state what should happen. A programmer converted that understanding into instructions. The computer executed those instructions quickly and reliably.
Consider inventory. “If stock falls below one hundred units, order five hundred more.” The machine can monitor every item continuously and place the order in milliseconds. But the threshold, quantity, objective, and relationship were supplied by people. The computer did not discover replenishment. It inherited a formalized decision.
Accounting software did not discover double-entry bookkeeping. Payroll software did not discover wages. Tax software did not discover tax law. Navigation software did not discover geometry. Spreadsheets did not discover arithmetic. Databases did not discover the meaning of a customer record.
These systems transformed civilization. Their lack of learning does not diminish them. It identifies their architecture.
Traditional programming is the transfer of understood patterns into executable form. The human says: “I know the transformation. I will describe it precisely enough that the machine can repeat it.”
This is why requirements were sacred. What should happen? Under what conditions? Which exceptions apply? How should the output be calculated? Someone had to know. Ambiguity upstream became failure downstream because the computer could only execute the formal structure it received.
The great constraint was not speed. It was articulation. Many tasks remained outside software because humans could perform them without being able to specify the complete rule.
Recognizing a cat is easy for a child and punishing for a rule writer. Four legs do not define a cat. Fur does not. Whiskers do not. Pointed ears do not. Every candidate rule admits other animals and rejects unusual cats. Human perception carried a pattern that human language could not finish describing.
Previous information technology could automate what we could formalize. The border of software was the border of explicit human understanding.
CH / 06
The inversion
Machine learning reverses the location of the pattern.
Instead of giving the machine a complete rule, we give it examples, a structure capable of changing, and an objective that distinguishes better predictions from worse ones.
The system predicts. The prediction meets an observed target. The difference becomes error. An optimization process changes parameters in a direction expected to reduce future error. Then the system predicts again.
Prediction. Error. Adjustment.
Repeated across enough examples, the system acquires internal structure that supports better performance. No programmer writes every feature, interaction, exception, or threshold. The programmer designs the learning arrangement: architecture, data pipeline, objective, optimization process, evaluation, and constraints. The particular predictive relationships emerge through training.
θ* = arg minθ 𝔼(x,y)~D[ℓ(fθ(x), y)]
Choose parameters θ that minimize expected loss ℓ over examples drawn from distribution D.The equation is compact. Its consequences are not. The objective does not state the rule for recognizing the cat, translating the sentence, detecting the fraud, or completing the proof. It states what counts as being wrong. Error supplies pressure. Training shapes a model that becomes less wrong on the data distribution.
This is the historical inversion:
Machine learning: Here are the examples. Find structure that predicts them.
Human understanding does not disappear. It moves. Humans still choose what to measure, what to optimize, what data to collect, what failures matter, what behavior is acceptable, and whether the output deserves use. But the human no longer has to enumerate the complete transformation before the machine can perform it.
Artificial intelligence is different from earlier information technology because useful structure can now emerge inside computation instead of entering computation only after a human has understood it.
Part III / The representation
The geometry of intelligence
A learned model does not merely store answers. It builds spaces in which difficult relationships become easier to express.
CH / 07
The geometry of intelligence
Representation is the hidden revolution inside machine learning.
The raw world does not arrive in a form that makes every useful relationship obvious. Pixels are not objects. Sound-pressure samples are not words. Tokens are not arguments. A strong model transforms inputs into internal representations where relevant distinctions can be made more easily.
Modern neural networks represent information as vectors: ordered lists of numbers. A word, image region, audio segment, or system state becomes a point in a high-dimensional space. Layers transform those points. Directions and distances acquire functional meaning because they help the system reduce predictive error.
The model is not required to label each direction in human language. It only needs the geometry to work.
Imagine that two classes overlap when plotted by height alone. Add width and they begin to separate. Add texture, motion, context, and history, and a boundary that was impossible in one coordinate system may become simple in another. Learning a representation means learning which transformations expose the structure needed by the task.
This is why deep learning broke through problems that had resisted hand-built features. The earlier system depended on humans to decide what measurements mattered. The learned system can transform raw inputs through many stages, developing features because they improve the objective.
h0 = x ; hk+1 = φ(Wkhk + bk)
Each layer transforms one representation into another. Training adjusts the transformations.The phrase “the model sees” is dangerous if taken literally. It invites a tiny human inside the network. A better statement is that the model maps inputs into a geometry whose organization supports prediction.
Humans also depend on representation. A map makes distance inspectable. Algebra makes unknown quantities manipulable. Musical notation makes temporal structure visible. A graph can reveal a relationship hidden in a table. Intelligence is not only possessing facts. It is finding a form in which the relationship can be used.
Machine learning expands that principle beyond forms humans can comfortably inspect. It can operate in thousands of dimensions without needing to picture them. This does not guarantee discovery. It creates representational room in which more complicated regularities can exist.
CH / 08
Dimensions are not parameters
A model can represent an item in a space thousands of dimensions wide while containing billions of learned parameters. These numbers describe different things.
A dimension is a coordinate in a representation. If a hidden state is a vector of 4,096 numbers, that state occupies a 4,096-dimensional space. The dimensions give the representation degrees of freedom.
A parameter is a learned numerical value that helps transform one representation into another. Weight matrices, biases, and related learned quantities contain parameters. Training changes them. Their collective values determine the function the network computes.
An activation is a value produced for a particular input as it moves through the model. Parameters persist across inputs. Activations change with the input.
Dimensions
Where a representation can vary. They describe the width and structure of a state space.
Parameters
The numerical values adjusted during training. They determine how representations are transformed.
Activations
The temporary values produced when one specific input passes through the learned machinery.
Confusing dimensions with parameters produces bad intuition. A billion-parameter model does not necessarily place each concept on its own billion-axis coordinate grid. Concepts are usually distributed across many activations and transformations. A single parameter rarely means “cat,” “democracy,” or “sarcasm.” The capability belongs to the organization of many values.
Open the distinction with a small network
Suppose a layer accepts a 100-dimensional vector and returns a 40-dimensional vector. Its weight matrix contains 40 × 100 = 4,000 parameters, plus 40 bias parameters. The output representation has 40 dimensions. The layer has 4,040 learned parameters. For each input, it produces 40 activation values.
Dimensions describe the input and output spaces. Parameters describe the transformation. Activations are the particular point produced this time.
Scale this structure across many layers and the numbers become enormous, but the distinction holds. Representation provides space. Parameters shape the terrain. Activations trace a route through it.
High dimensionality is not magic. More room can support richer structure, but it can also support noise, spurious shortcuts, and brittle boundaries. Capacity becomes intelligence only when training, data, and evaluation force the capacity toward structure that generalizes.
CH / 09
Random to whom?
Humans often use the word random when the more accurate sentence is: “I cannot identify a representation that predicts this.”
Those claims are not equivalent.
Consider XOR, the exclusive-or relation. Two binary inputs produce 1 when they differ and 0 when they match. Examine either input by itself and the output remains perfectly balanced. Neither variable alone predicts the answer. Examine both together and the answer is deterministic.
Inspect the XOR field
Projection destroys the relationship. In the joint representation, the pattern is exact.
XOR is elementary. Its lesson is not. A pattern can vanish when the observer projects the data into too few dimensions. Nothing about the underlying process changed. Only the representation changed.
Now replace two binary inputs with hundreds of interacting variables. No single measurement may carry much signal. Pairs may look weak. Simple charts may show noise. Yet a high-order relationship could still reduce uncertainty when the variables are represented together.
This is where machine learning can exceed normal human pattern discovery. Human conscious reasoning is powerful but representationally narrow. We favor a small number of named variables, visible trends, and relations that can be verbalized. A trained model can exploit a distributed relationship across more coordinates than a person can hold in explicit attention.
That does not abolish randomness. Quantum measurement, thermal noise, unobserved causes, chaotic sensitivity, and computational irreducibility all place different limits on prediction. Even a deterministic process can become practically unpredictable when tiny errors in initial conditions grow rapidly.
The correction is narrower and more consequential:
The arrival of stronger pattern-finding systems therefore changes the status of the unknown. “Unpredictable” can no longer be treated as one category. We must ask whether the barrier is fundamental, informational, representational, computational, or temporary.
Part IV / The epistemic machine
Useful before understood
When a machine finds predictive structure first, human knowledge no longer has to precede every useful computation.
CH / 10
Useful before understandable
Under the old order, understanding usually preceded usefulness. A human discovered a relationship, converted it into a method, and then asked a computer to execute it.
Machine learning permits the opposite order. A model can exploit a relationship that improves prediction before a human can express that relationship as a compact rule.
Give a system rich patient data and verified outcomes. Ask it to estimate the probability of a disease. It may discover that a complicated interaction among measurements, history, timing, and image features changes risk. The prediction may be reproducibly useful even when no physician can summarize the interaction in one sentence.
This creates a historically unusual conversation.
The machine predicts correctly.
The human asks: “Why?”
With traditional software, the answer was available in principle because the machine knew only the logic we supplied. With a learned model, the effective logic is distributed across parameters and conditional activations shaped by training. The model may be inspectable in fragments without yielding a complete human-scale explanation.
Predictive accuracy asks whether the output is right. Interpretability asks whether humans can trace relevant internal or input-output relationships. Explanation asks whether we possess a satisfying account of why the phenomenon itself occurs. These standards can move together. They are not identical.
The existence of a useful black box does not eliminate the need for explanation. In medicine, law, finance, and public administration, the reason can matter as much as the score. A prediction can be accurate on average and unjust in application. It can rely on a proxy that fails after a policy change. It can expose a correlation that is real but unsafe to manipulate.
The new possibility is not “accuracy replaces understanding.” It is “accuracy can arrive before understanding.”
That reversal gives human inquiry a new starting point. Instead of asking only, “What theory should we encode?” we can ask, “What structure is the model exploiting, where does it hold, and can we turn it into a testable human account?”
The model becomes an instrument for detecting the presence of knowable structure. It does not automatically deliver the knowledge in human form.
CH / 11
Computation before understanding
For centuries, formal computation followed conceptual discovery.
Newton supplied equations; people calculated their consequences. Engineers developed theories of load; computers solved larger structures. Accountants established financial categories; software produced statements. A human theory defined the meaningful variables and relationships before computation scaled them.
Artificial intelligence permits another sequence:
The machine can provide evidence that structure exists before a human knows its form. A stable predictive gain is not yet a scientific explanation, but it is a signal: some information in these observations reaches the outcome.
This changes computation from a calculator of known relationships into a detector of possible relationships.
The difference is as profound as the difference between a telescope and a table of planetary positions. The table stores known observations. The telescope extends observation. A learning system extends pattern detection. It can make regularity operationally visible through improved prediction even when the human eye sees only a dense field of variables.
But computation before understanding introduces a new discipline. Predictive performance must survive outside the data that produced it. Training error is not discovery. A flexible model can memorize. A leaky dataset can reveal the answer through an accidental channel. A benchmark can reward the wrong shortcut. Repeated testing can manufacture significance.
The evidence becomes serious only when the relationship survives appropriate separation: new samples, new times, new settings, adversarial checks, changed measurement procedures, and competing hypotheses.
Computation can precede understanding. It cannot excuse validation.
Nor can it settle ontology by itself. A model may use a variable because it predicts. That does not tell us whether the variable is a cause, a consequence, a proxy, a measurement artifact, or a marker for something omitted. Predictive structure opens the investigation. It does not close it.
CH / 12
From automation to discovery
Automation removes execution from a known transformation. Discovery searches for a transformation that was not already known.
A factory robot automates a motion. Payroll software automates a calculation. A database automates storage and retrieval. A compiler automates translation between formal representations. Each system can be remarkably complex. In each case, people specify the operation and its success conditions.
Machine learning can participate earlier. It can search over functions, representations, designs, molecules, policies, proofs, or control strategies and use predictive feedback to move toward useful regions.
This is not unrestricted discovery. Every search has a space, an objective, evidence, and constraints. The model can only find what the setup makes expressible and what the feedback makes selectable.
The cleanest way to see the difference is:
Automation
Humans specify the rule. The machine repeats it faster, cheaper, or more reliably.
Learning
Humans specify examples and loss. The machine fits structure that generalizes.
Discovery
Prediction and search expose candidates that humans did not enumerate in advance.
Discovery emerges when the learned structure supports a candidate that changes the human search. A model proposes a molecule with desired properties. It identifies a mathematical relation. It reveals an unexpected subgroup. It generates a design that satisfies constraints. The candidate remains a prediction until reality answers.
That final sentence preserves the boundary. AI predicts. Experiments meet the world. Institutions decide what evidence counts. Humans accept responsibility for acting.
A generated hypothesis is not a discovery merely because it is novel. A predicted protein structure is not a treatment. A plausible proof is not a theorem until verified. A high-scoring policy is not legitimate until its consequences and values are examined.
Artificial intelligence accelerates the production and ranking of possibilities. It changes discovery by making search over rich spaces cheaper. But discovery still requires the conversion of a predicted possibility into an actual, tested relationship.
CH / 13
When noise becomes information
Information is not simply data. It is a reduction in uncertainty relative to a question.
A laboratory archive may contain millions of measurements that nobody can use. The data exists. The information becomes visible when a representation links part of that archive to an outcome, a mechanism, or a decision.
Machine learning changes the economics of making those links. It can test more interactions, absorb more examples, and form richer intermediate representations than a person can inspect manually. As predictive capability rises, some observations move from the category “noise” into the category “signal.”
This movement is never automatic. A larger model can also turn accidents into apparent patterns. If the dataset records the hospital scanner rather than the disease, the machine may predict the source institution. If the answer leaks into a timestamp, the model may appear brilliant. If historical discrimination shaped the target, predictive accuracy may reproduce the discrimination.
There are therefore two opposite errors:
- Human provincialism: assuming no pattern exists because humans cannot see one.
- Machine credulity: assuming every model gain reveals durable structure in the world.
Advanced students must refuse both.
The correct question is not “Did the model find a pattern?” Flexible models always find something. The correct questions are: Does it survive new data? Does it survive intervention? Does it survive measurement change? Does it transfer across populations? Does the gain exceed a simpler baseline? Does the representation rely on information that will exist when the prediction is needed?
When those tests succeed, the status of the data changes. What was formerly too entangled to use becomes predictive. The archive becomes an instrument.
This is one reason AI alters the boundary of economically useful information. A variable can matter even when no analyst would have chosen it in advance. A weak trace can matter in combination. Old records can acquire new value when better representation makes their structure accessible.
The world did not become less structured. The observer became more capable.
CH / 14
The revised scientific method
Artificial intelligence does not replace the scientific method. It inserts a new instrument between observation and explanation.
The classical ideal moves from observation to hypothesis, from hypothesis to experiment, and from experiment to revised theory. The AI-assisted path can begin with predictive structure that no human proposed:
The model identifies where prediction improves, which cases violate expectation, which variables interact, or which candidate deserves testing. Humans then design probes that distinguish among possible explanations.
This suggests a revised division of labor.
- Machines search representational space. They fit patterns across scales and combinations that exceed unaided inspection.
- Humans formulate epistemic tests. They ask which alternative explanations remain and what intervention would separate them.
- Reality adjudicates. The experiment, deployment, or new observation determines whether the predicted structure persists.
- Institutions govern use. They decide when evidence is sufficient, which harms are unacceptable, and who is accountable.
The model can also become the object of science. Mechanistic interpretability, feature visualization, attribution, sparse representations, causal tracing, and controlled ablation all attempt to convert learned computation into inspectable claims. The goal is not always to translate the entire network into a paragraph. It is to establish specific, testable statements about what information a system uses and how that use affects output.
Science advances when the cost of a meaningful question falls. AI can lower the cost of proposing hypotheses, screening candidates, finding anomalies, and designing measurements. That increases the importance of standards. Cheap hypotheses can flood a field. Cheap correlations can exhaust attention. The scarce resource becomes discriminating evidence.
CH / 15
What AI still cannot predict
No pattern, no prediction. No relevant observation, no access to the pattern. No stable environment, no guarantee that yesterday’s pattern survives tomorrow.
The limits of prediction are not one wall. They are different walls that require different responses.
Irreducible uncertainty
Some events may contain genuine stochasticity. Even with a perfect description of the available state, outcomes retain a distribution. A good model can estimate that distribution. It cannot convert probability into certainty.
Missing state
A process can be patterned while the predictor lacks the variables that carry the pattern. The limitation is observational. More data of the same kind may not help. A different sensor might.
Chaotic sensitivity
Deterministic systems can amplify microscopic uncertainty. Weather has structure and remains bounded in forecast horizon. Better models extend the horizon; they do not abolish sensitivity to initial conditions.
Distribution shift
A model learns from one distribution and meets another. Technology changes, language moves, pathogens evolve, incentives adapt, and institutions rewrite rules. A pattern can be real and expired.
Reflexivity
Predictions can change the system they predict. A public market forecast affects trades. A policing score affects where police look. A ranking changes what people click. The model enters the causal field and alters the data-generating process.
Computational limits
A pattern may exist but require infeasible computation to exploit. The shortest description can be impossible to find. The best plan can be beyond practical search. Solomonoff induction is an illuminating ideal precisely because it is not a usable finite procedure in the general case.
Objective failure
The model can accurately predict the target we supplied while the target fails to represent what we value. Loss functions are not moral theories. Benchmark victory is not social legitimacy.
Open the prediction-limit audit
Before calling a system unpredictable, ask: Is the uncertainty intrinsic? Are the relevant variables observed? Is the representation adequate? Is the process stable? Does the prediction change behavior? Is the computation tractable? Is the objective the real objective?
Each answer points toward a different action: accept a distribution, collect a variable, learn a representation, monitor drift, model feedback, approximate the search, or redefine success.
The mature claim is not that AI predicts everything. It is that AI expands the class of patterns that can become predictively useful before humans can formalize them. Expansion is not infinity.
CH / 16
The boundary of the knowable
Artificial intelligence changes the boundary between what humans understand and what human systems can use.
Before machine learning, the usable boundary sat close to the articulable boundary. To make software perform a transformation, humans generally had to express the transformation. The machine extended speed, scale, memory, and reliability. It did not usually extend the discovery of the operative pattern.
Now the boundaries separate.
A relationship may be unusable by unaided humans, usable by a model, partially interpretable by investigators, and still not understood as a compact theory. Use can precede explanation. Prediction can precede science. Computation can reveal the outline of a regularity before language catches it.
This creates at least four domains:
- Known and formalized: humans understand the rule and machines execute it.
- Learned and interpretable: machines help discover structure that humans can subsequently state and test.
- Learned and operational: prediction works within validated bounds, but the full internal relationship resists human summary.
- Unreachable: relevant structure is absent, unobserved, unstable, or computationally inaccessible.
The third domain is the new tension. It offers enormous value and demands enormous care. A society can act on relationships it cannot fully explain. Sometimes that is rational. Humans have always used medicines, heuristics, and institutions before possessing complete theories. But machine learning scales the practice and obscures the path by which the relationship was formed.
Responsibility therefore becomes more important, not less. The inability to explain a model does not transfer moral agency to the model. People choose the target, accept the evidence, define the deployment, distribute the errors, and decide who may appeal.
Prediction enlarges capability. It does not answer what should be done with the capability.
The boundary of the knowable is no longer a single line between ignorance and understanding. It is a layered frontier among what can be predicted, what can be compressed, what can be explained, what can be caused, what can be controlled, and what should be used.
Artificial intelligence moves those lines at different speeds.
Coda / The difference
We no longer have to arrive first.
Previous information technology waited for human understanding. We found the pattern. We stated the rule. We built the program. The machine executed our discovery.
Artificial intelligence changes the sequence. We can give the machine experience, an objective, and a way to change. Predictive error can shape a representation. The representation can exploit structure that no person enumerated. The pattern can become useful before it becomes humanly legible.
That is why AI is not simply faster software, friendlier software, or software that speaks. Those are surface descriptions.
The deep difference is epistemic.
For the first time at scale, information technology can participate in finding the transformation instead of only executing a transformation already found.
This does not make the machine omniscient. It does not erase randomness, causation, experiment, judgment, or responsibility. It does something more exact.
It breaks the old dependency.
The machine no longer needs us to find the pattern first.
Sources and further coordinates
- Claude E. Shannon, “A Mathematical Theory of Communication”, 1948. The foundation for entropy, information, and probabilistic code length.
- Ray J. Solomonoff, “A Formal Theory of Inductive Inference”, 1964. A foundational union of algorithmic description, probability, and prediction.
- Shane Legg and Marcus Hutter, “Universal Intelligence: A Definition of Machine Intelligence”, 2007. A formal treatment of intelligence as goal achievement across environments.
- Yoshua Bengio, Aaron Courville, and Pascal Vincent, “Representation Learning: A Review and New Perspectives”, 2012. A comprehensive account of learned representations.
- Grégoire Delétang et al., “Language Modeling Is Compression”, 2023. Experimental and operational treatment of the prediction-compression equivalence.
- Yuzhen Huang et al., “Compression Represents Intelligence Linearly”, 2024. Empirical evidence relating compression performance to benchmark capability, with scope limitations.
- David H. Wolpert and William G. Macready, “No Free Lunch Theorems for Optimization”, 1997. The formal reminder that performance depends on assumptions about problem structure.
1 thought on “We Used to Find the Pattern First”