1 Facing the Unknown: How Does Humanity Seek Answers?

Humans have always lived in a world filled with uncertainty: whether in the natural environment, social relationships, or personal choices, we cannot directly know what has not yet occurred or what remains undiscovered. Therefore, how to use the information already mastered to understand the unknown is a fundamental cognitive capacity that humanity has developed over time in the face of an uncertain world.

This capability is ubiquitous in modern society: financial institutions analyze market changes through historical data, enterprises determine consumer demand based on user behavior, meteorological systems use complex models to infer weather, and medical research evaluates disease risks through statistical analysis. With the advent of artificial intelligence, machines have also begun learning patterns from massive amounts of data, identifying correlations between information, and applying these capabilities to tasks such as image recognition, speech processing, and text generation, as well as inferring unknown outcomes.

However, humanity's exploration of the unknown did not begin with modern science. In ancient times, lacking statistics, computers, and modern scientific tools, people faced many questions that had no straightforward answers: How will my circumstances turn out in the near future? Is this person worthy of trusting with my life? Faced with major life choices, what results might different decisions bring? To cope with these uncertainties, people also searched for grounds to help them make judgments.

Consequently, various methods for attempting to understand the unknown gradually developed throughout history, including traditional systems such as divination, astrology, and destiny analysis. Although built on entirely different knowledge bases and employing completely distinct modes of explanation and bases of judgment compared to modern science and artificial intelligence, they can all be viewed as different attempts by humanity to construct systems of interpretation and judgment for the unknown.

This raises a question worth pondering: If ancient fortune-telling, modern science, and artificial intelligence all attempt to understand the unknown, what is the essential difference between them? And how did humanity step by step develop this exploration—initially reliant on experience and symbols—into the complex inference made possible today by data, models, and computational power?

2 How the Ancients Inferred the Unknown from the Known

1 How Did the Ancients Understand and Infer the Unknown?

Humanity has not had only one way of attempting to understand the unknown. Across different civilizations and historical eras, people developed diverse methods of explaining and judging the unknown based on the phenomena observable at the time, accumulated experience, and their understanding of the world.

These methods differ vastly in form: some infer trends of change by observing natural phenomena, such as the ancients judging potential changes through celestial patterns, climate, and natural laws; some infer life trajectories through personal information, such as establishing destiny analysis systems based on birth time; still others trigger reflection through random symbols, such as using hexagrams or cards to help people re-examine their current situation.

Viewed through the lens of information processing and model building, these systems, though utilizing different symbols, variables, and explanatory frameworks, share certain formal similarities: they all attempt to acquire information from reality or specific contexts, reduce complex problems into a finite set of symbols or variables, and then form judgments about the unknown according to their respective rules of interpretation.

Of course, formal similarities do not imply that these systems possess the same reliability or epistemological foundation. Modern science relies on systematic data collection, mathematical methods, and replicable verification, whereas ancient systems relied more on empirical accumulation, symbolic systems, and traditional explanatory frameworks.

Nevertheless, from the perspective of human cognitive development, these ancient systems remain of great significance: they represent an early attempt by humanity, in the absence of modern scientific tools, to observe, generalize, and explain a complex world.

2 I Ching Divination: Understanding the Unknown Through Change

Among the numerous ancient methods for judging the unknown, the I Ching stands as a prominent representative of systematic thinking.

The concept of “Yi” in the I Ching is commonly associated with change. The ancients believed that everything between heaven and earth is in constant flux, that nothing remains in an entirely fixed state, but instead transitions perpetually between different stages. Therefore, the question the I Ching focuses on is not merely: What will definitely happen in the future? Rather, it is closer to: What state is the current situation in? Toward what might this state evolve? How should a person adjust their actions within this process of change? To understand how the I Ching describes such change, one must first understand the fundamental symbolic system it employs.

The I Ching describes the diverse attributes and states of things through the fundamental concepts of Yin and Yang. A Yao (line) is the basic symbol representing Yin and Yang in the I Ching, divided into Yin lines and Yang lines, represented respectively by a broken line “⚋” and an unbroken line “⚊”. These two primary forms are referred to as the “Two Modes” (Liangyi).

On this basis, combining two lines in different ways produces four structures, termed the “Four Images” (Sixiang, 2²=4); further combining three lines yields eight fundamental structures, known as the “Eight Trigrams” (Bagua, 2³=8):

Further stacking two three-line trigrams on top of each other forms a six-line hexagram. Because each line still has only two possibilities—Yin or Yang—the six lines yield a total of 2⁶=64 possible combinations, which forms the system of the Sixty-Four Hexagrams:

In the anime Naruto, Neji's ultimate technique also borrowed from this concept:

Re-examining this way of thinking from the perspective of modern systems theory, one can understand it as a description of system states and their evolutionary processes. In modern terms, this can be understood as the state evolution process of a system from dormancy, through development, to maturity. This staged structure finds certain parallels in modern corporate lifecycles, technology adoption cycles, and even personal growth models.

For instance, the lines in the Qian (The Creative) hexagram—“Hidden dragon, do not act,” “Dragon appearing in the field,” and “Flying dragon in the heavens”—do not describe three entirely distinct lives, but rather a development process in which capability and environment continuously align: “Hidden dragon, do not act” emphasizes that although one possesses potential, conditions are not yet ripe in the current stage, making it unsuitable to rush into action; “Dragon appearing in the field” signifies that abilities are beginning to show and one is gradually entering an environment where they can be exercised; “Flying dragon in the heavens” represents a state where personal capability, environmental conditions, and the opportunities of the times have reached a high degree of harmony.

Therefore, from a modern viewpoint, the significance of I Ching divination does not necessarily lie in whether its judgment of a specific outcome is accurate, but in the fact that it provides a framework for thinking in the face of uncertainty: when confronting major choices, people tend to focus directly on “what will the outcome be,” whereas the I Ching focuses more on what state one is currently in, what changes may take place, and what actions one can take in response to those changes. From this perspective, it concerns not only the outcome itself, but also the process of change prior to the outcome taking shape.

Abstracted from the perspective of model construction, the basic logic of I Ching divination can be summarized as follows:

现实问题
    ↓
随机产生卦象
    ↓
阴阳变化模型
    ↓
分析当前状态与发展趋势
    ↓
形成行动建议

In actual divination practice, one common method is the coin toss method (also called the Six-Yao coin divination): three copper coins are tossed simultaneously each time, and the combination of heads and tails determines the Yin or Yang attribute of a line, as well as whether a change occurs. The various combinations of three coins correspond to four outcomes: Old Yin, Young Yin, Young Yang, and Old Yang, among which Old Yin and Old Yang are changing lines. After six consecutive tosses, a complete six-line hexagram is formed from bottom to top, which is then interpreted in conjunction with the hexagram statement, line statements, and line change relationships; when necessary, a new hexagram (the transformed hexagram) is generated based on the changing lines.

Understood from a modern cognitive perspective, randomly generated hexagrams can be regarded as a trigger mechanism for thinking: by introducing non-predetermined information, they help people step out of their preexisting cognitive frameworks and re-examine the problem at hand. Of course, the hexagram itself cannot directly provide answers; it must be coupled with the underlying philosophy of Yin-Yang transformation and interpretive systems like the hexagram statements and line statements to formulate a judgment on real-world issues.

From a formal perspective, the symbolic system of the *I Ching* also shares similarities with the concept of binary in modern information encoding: yin and yang lines constitute two fundamental states, and through combinations of a limited number of binary states, more complex information structures can be expressed. Binary uses 0 and 1 to represent two basic states and expresses more complex data through different combinations; its core combinatorial thinking bears a certain resemblance to how yin and yang lines compose hexagrams.

Therefore, rather than being a tool for predicting the future, the *I Ching* is better understood as a system of state description and interpretation established by ancient humans when confronting complex changes. What it seeks to answer is not “what will definitely happen in the future,” but rather “in an ever-changing environment, where do I stand now, and how should I act?”

3 BaZi and Zi Wei Dou Shu: Understanding the Individual from Limited Information

If the *I Ching* focuses more on state transitions in the process of change, then BaZi and Zi Wei Dou Shu further shift their focus toward the individual: can a person's personality, circumstances, and life experiences be explained starting from innate information such as their time of birth?

The common logic behind these destiny-calculation systems is to start from limited information such as birth time, convert it into specific symbols and attributes, and then, through a set of interpretation rules, connect this information with human personality, experiences, and life changes.

Among them, BaZi is one of the most representative traditional systems of destiny calculation. BaZi, also known as the “Four Pillars,” uses a person's birth year, month, day, and hour as basic information, converting them into the Year Pillar, Month Pillar, Day Pillar, and Hour Pillar. Each pillar consists of a Heavenly Stem and an Earthly Branch, thus totaling eight characters. For instance, for a male born in Chengdu at 12:30 PM on December 1, 1980, after converting his birth time according to traditional BaZi chart-casting rules, the following Four Pillars can be obtained:

From a modern perspective, BaZi actually accomplishes a process of information encoding: birth time is converted into a set of symbolic variables with specific meanings, which are then analyzed combinatorially through rules such as the generation and overcoming of the Five Elements and Yin-Yang attributes, ultimately forming explanations of personal character, aptitude tendencies, and life stages. Its structure can be abstracted as:

出生信息
    ↓
符号编码（天干地支）
    ↓
属性分析（阴阳五行）
    ↓
个人特征与发展趋势判断

Zi Wei Dou Shu is a further expansion based on similar thinking. It likewise uses birth time as foundational input, but introduces more structured variables, such as the Twelve Palaces and various stellar combinations, to describe life domains such as career, wealth, family, and interpersonal relationships. Compared to BaZi, Zi Wei Dou Shu attempts to form a holistic interpretation of an individual and their life stages through more dimensions and their combined relationships:

For instance, in the traditional interpretive system, the Life Palace is primarily used to describe personal character and fundamental traits, the Career Palace is used to analyze career development, and the Wealth Palace analyzes financial circumstances. By observing the star placements in different palaces and their interrelationships, Zi Wei Dou Shu seeks to construct a multifaceted, holistic understanding of a person:

出生信息
↓
转换为多个属性维度
↓
分析不同维度之间的关系
↓
形成对个人及人生阶段的整体解释

From this perspective, BaZi and Zi Wei Dou Shu reflect a typical mode of human cognition: when faced with a complex life system, people seek to define a few key variables first, and then use these variables to further explain a vast array of complex phenomena.

Of course, from the perspective of modern science, there is currently no reliable evidence proving that a person's birth time can accurately predict their character and life trajectory through the rules adopted by BaZi or Zi Wei Dou Shu; thus, these systems cannot be regarded as scientifically validated predictive models. However, from the standpoint of cognitive development, they indeed reflect an early human attempt to transform complex, uncertain problems into comprehensible structures.

4 Tarot Cards: Random Symbols and Meaning Interpretation

Unlike the I Ching, BaZi, and physiognomy, Tarot originated in Europe, initially used primarily as a card game, and later gradually evolved uses such as divination and symbolic interpretation. Modern conventional Tarot decks usually consist of 78 cards, comprising 22 Major Arcana cards and 56 Minor Arcana cards:

Among them, the Major Arcana is primarily used to express broader, macroscopic life themes, such as: The Fool: beginnings, exploration; The Magician: creation, action; Death: endings, transformation; The World: completion, integration. The Minor Arcana, on the other hand, is more similar in structure to standard playing cards, divided into four suits—Wands, Cups, Swords, and Pentacles—each corresponding to different symbolic themes, and is more often used to describe specific situations in daily life.

Tarot does not have a single, fixed interpretive system; different traditions adopt varying card designs, symbolic origins, and interpretive methods, such as the classic Rider-Waite-Smith, Marseilles, and Thoth systems. Although these systems share some traditional symbols, they differ in card symbolism, spreads, and methods of interpretation. Therefore, Tarot is not a fixed predictive model governed by unique interpretive rules, but rather an open-ended symbolic language. The same card can yield different interpretations under different systems and specific contexts.

Tarot also differs from the aforementioned systems in its source of information and method of judgment: the I Ching generates hexagrams through random methods, BaZi and Zi Wei Dou Shu establish grounds for judgment based on birth information, while Tarot randomly draws symbolic signs, allowing users to interpret meaning centered around these symbols.

This process can be abstracted as:

随机符号
    ↓
象征解释
    ↓
个人经验投射
    ↓
认知调整与行动思考

The randomly generated cards do not inherently contain objective information about an individual's future, but they provide a set of symbolically meaningful stimuli that allow people to reorganize their own thoughts around these symbols. For example, the “Death” card in Tarot is often misunderstood as representing death or disaster, but in most interpretive systems, it symbolizes endings, transformation, and new beginnings. When a person confronts such a symbol, what truly takes effect may not be any information the card provides about the future, but rather how its symbolic meaning prompts the user to re-examine their current state and reflect on whether they need to alter their existing behavioral patterns.

From a psychological perspective, this process shares certain similarities with human projection mechanisms: when faced with external symbols that hold ample room for multiple interpretations, people often project specific meaning onto them by drawing from their own experiences, emotions, and current concerns. Therefore, in a sense, Tarot functions more like a cognitive mirror:The randomly generated symbols provide an external stimulus, while the process of interpreting these symbols may help individuals uncover thoughts they were not previously consciously aware of.

Of course, this does not mean that Tarot can accurately predict specific events; from the perspective of modern science, there is currently no reliable evidence proving a stable causal relationship between randomly drawn cards and concrete occurrences. Yet, viewed through how humans navigate the unknown, Tarot embodies an alternative approach: when unable to obtain direct answers, one can use external symbols to re-evaluate their situation, organize existing perceptions, and assist themselves in making judgments.

5 Facing the Unknown: What Does Humanity Need?

Although the I Ching, Bazi, Zi Wei Dou Shu, and Tarot cards employ different symbolic systems and interpretive methods, their enduring popularity stems not only from people's desire to predict the future, but also from how they satisfy fundamental human cognitive needs when facing an uncertain world.

One of their vital functions is reducing the sense of uncertainty brought by the unknown. When confronting major life dilemmas such as marriage, career, illness, and wealth, people often seek explanations that make the otherwise incomprehensible open to discussion, or even actionable. What these systems provide is precisely a framework for understanding and interpreting one's current situation: while it may not truly unveil the unknown, it helps people form relatively clear perspectives and expectations.

Another function is providing a psychological reference point for decision-making. When facing critical choices, people do not always possess sufficient information, nor can they always rely entirely on rational calculations. Whether it is a hexagram, an astrological chart, or a tarot card, each can serve as an external reference that prompts one to re-examine their circumstances. Sometimes, what a person truly needs is not an absolute, correct answer, but a benchmark that helps them organize their thoughts, confirm their inclinations, or cope with uncertainty.

This also explains why such needs have not vanished with the advancement of science—in modern society, people still rely on tools like weather forecasts, economic projections, risk models, and big data analytics. Although their data foundations, theoretical bases, and validation methods are fundamentally different from ancient divination and astrological systems, they share a similar underlying human drive: the desire to use known information to understand the unknown, and to make judgments and take action accordingly.

3 Why Can Experience Serve as the Foundation for Inferring the Unknown?

1 From Observation to Experience: How Humans Accumulate Knowledge of the World

Chapter Two discussed how the ancients established frameworks of interpretation and judgment for the unknown through various means, such as hexagrams, birth details, or random card draws; however, these rules for interpretation and judgment did not appear out of thin air. How, then, did people gradually accumulate the bases for these judgments?

In fact, one of humanity's earliest ways of understanding the world was through long-term observation and empirical accumulation: people in agrarian societies observed seasons, weather, and crop cycles; physicians observed the relationships between symptoms and illnesses; merchants tracked fluctuations in prices, supply, and demand; and military strategists drew lessons on how terrain, troop numbers, and tactics influenced the outcomes of repeated battles.

Initially, such experiences may have been merely scattered judgments: “When this happens, that usually follows.” Yet as similar phenomena recurred, people attempted to identify patterns and distill these observations into rules of thumb that were easier to remember and pass down. With continuous accumulation, practice, and transmission, certain frequently observed phenomena and their correlations gradually evolved from isolated individual insights into collective, time-tested knowledge.

Thus, while antiquity lacked databases, sensors, and computers, humanity itself functioned as an enduring “observation system.” The fragmented experiences accumulated by countless individuals through long-term practice, though devoid of modern data collection and structured processing, nonetheless formed a vast repository of historical experience—viewed through today's data mindset, it resembles an “ancient big data” without technical infrastructure.

This is where the question becomes intriguing:If an empirical rule is indeed backed by a large volume of historical samples, just how different is it from what we call data analysis today?

2 From Experience to Patterns: The Case of “Darkness Between the Eyebrows”

To understand the difference between ancient empirical wisdom and modern data analysis, a classic example is the traditional folk saying about “a darkened brow (yintang)”:

As this experience was continuously observed and passed down, the originally complex and ambiguous observations were likely compressed into a memorable, easily shareable maxim: “A darkened brow presages imminent misfortune.” What occurred here was effectively a classic form of information compression: complex personal conditions, behavioral manifestations, and subsequent outcomes were reduced to an easily recognizable external feature, which was then mapped directly to a specific outcome.

If redescribed in the language of modern data analysis, this process can be understood as a very rudimentary form of pattern recognition: observers search for recurring features across extensive historical experience, and then determine whether a certain outcome might occur based on these features. Modern machine learning also seeks statistical relationships between features and outcomes from historical samples, and then applies these relationships to new ones. The two share a fundamental similarity in their underlying logic: both attempt to identify patterns from past occurrences and use those patterns to understand the unknown.

However, when ancient people accumulated experience and sought patterns, they were easily influenced by cognitive biases. If something bad actually happened to someone with a “darkened forehead,” that instance was likely remembered and cited as evidence supporting the empirical rule; conversely, countless cases where nothing happened despite a “darkened forehead” might be overlooked because they lacked drama. As a result, what people remembered and passed down was no longer the full scope of observations, but rather those cases that better aligned with their preexisting judgments.

Therefore, even if an empirical rule is genuinely supported by an accumulation of extensive historical observations, one cannot assume its pattern is reliable simply because it has endured for a long time or is widely accepted. Discovering patterns from experience is merely the beginning of recognizing underlying laws; determining whether such patterns genuinely exist, can be reliably reproduced, and possess real explanatory and predictive power requires further verification.

3 From Empirical Patterns to Testable Laws

This brings us to a more critical question: if ancient people could accumulate experience through sustained observation and identify recurring patterns from it, why couldn't these patterns directly serve as reliable laws?

The true distinction lies not merely in the volume of data, but in the fact that humanity began employing fundamentally different methods to judge whether an observation qualifies as a law.

Lacking systematic analysis and verification tools, ancient empirical knowledge was often inclined to directly summarize repeatedly observed relationships into rules. For instance, observing that a specific trait frequently co-occurred with a certain outcome might produce the empirical judgment that “when A appears, B usually follows.” This method was simple, intuitive, and readily transmitted through language and stories, yet it rarely probed deeper: What is the probability of this relationship occurring? Are other factors concurrently influencing the outcome? If A does not occur, would B still happen? Does this relationship still hold across different populations and environments?

The development of modern statistics and scientific research methods precisely enabled people to pose these questions systematically. No longer content with merely stating “I observed a pattern,” people attempt—through probabilistic analysis, controlled studies, and replicated verification—to determine whether an observed relationship exceeds the scope of random fluctuation and whether it can consistently hold under varying conditions.

For example, if it is observed that “people with darkened foreheads are more prone to misfortune,” modern statistical thinking would not accept this conclusion outright, but would instead investigate further: Among those with darkened foreheads, how many actually encountered misfortune? Among those without this feature, how many encountered misfortune? Is there a statistically significant difference in rates between the two groups? Could this difference be confounded by other variables such as age, health status, or life stress?

In other words, ancient empirical knowledge primarily sought “patterns that appear to exist,” whereas modern statistics began attempting to determine “whether such patterns truly exist.”

This represents a profound cognitive leap: the former relies on long-term observation and empirical induction, while the latter demands that empirically induced patterns be rigorously tested through probability, controls, and repeated verification.

From this perspective, these ancient systems were not crude, rudimentary versions of modern science, but rather distinct early approaches through which humanity sought to understand and explain the unknown. Although they relied on different sources of information and explanatory frameworks, they all reflected, to varying degrees, the foundational concept of “starting from known information to infer the unknown through existing patterns.” They simply had not yet established a systematic methodology to distinguish whether the grounds for these inferences reflected genuine laws or merely coincidental correlations extracted from a complex reality.

4 From Empirical Judgment to Scientific Inference

1 From Empirical Patterns to Mathematical Models

The end of Chapter 3 discusses a crucial boundary: humanity must not only discover patterns from historical experience, but also determine whether these patterns genuinely exist and translate them into testable laws. Yet, even when this step is accomplished, it still does not directly produce a system ready for prediction. That is because between discovering a law and using it to calculate the future lies another indispensable link—translating that law into a computable model.

A model, simply put, can be understood as a structure established to describe variables and their relationships after abstracting and simplifying the real world. Variables in the real world are often extremely complex and mutually influential; therefore, humans need to extract variables relevant to the problem, describe their connections using certain mathematical relationships, and then use these relationships to infer states that may arise in the future.

Once these relationships are organized into a runnable model, prediction becomes a computable process:

现实世界
    ↓
观测数据
    ↓
识别变量及其关系
    ↓
数学模型
    ↓
输入当前状态
    ↓
计算未来状态

For instance, “a continuous drop in barometric pressure may indicate a changing weather system” can be considered an empirical rule of thumb, but on its own, it is insufficient for modern weather forecasting. True prediction also requires knowing what the current pressure is, how fast it is changing, the states of temperature and humidity, and how these variables jointly affect future weather. Only when these relationships are organized into a functioning model do patterns truly acquire predictive power.

This is also a crucial distinction between modern predictive models and traditional rules of thumb. Empirical rules often compress complex reality into a few easy-to-remember correspondences, whereas models attempt to retain more variables that affect outcomes and describe the interactions among them. The more complex reality is, the more variables and intricate interactions a model typically needs to account for.

Similar shifts have occurred in fields such as finance, medicine, engineering, and the social sciences: financial models forecast trends based on market variables, medical models assess disease risks based on patient data, and engineering models predict future changes based on equipment status. Although their subjects of study differ, they all follow the same rationale: organizing patterns of the real world into a model that accepts inputs and produces outputs.

Therefore, from the perspective of modern science, “discovering patterns” and “building models” are two interconnected yet distinct stages. The former addresses whether we have identified a stable relationship, while the latter addresses how to transform that relationship into a computable structure and use it to make predictions for new situations.

2 From Determinism to Probability

Once humans began using models to forecast the future, they soon encountered an unavoidable problem: even if a model is founded on stable patterns, it is still difficult to guarantee that the predicted outcome will definitely occur. The real world contains a vast number of factors that cannot be fully observed, precisely controlled, or are inherently random; hence, even if a model can describe the underlying patterns, it can hardly guarantee that the predicted outcome will inevitably materialize.

For example, a weather model can predict whether it might rain tomorrow based on the current atmospheric state, but even with numerous variables considered, it is hard to simply provide a definitive conclusion like “it will definitely rain tomorrow.” A more reasonable way of expression is to provide the probability of precipitation within a specific time and area. Therefore, modern weather forecasts usually express forecast results in terms such as “probability of precipitation.”

This signifies a major difference between modern prediction and judgments based on traditional experience. Traditional empirical rules tend to establish relatively direct correspondences, whereas modern predictive models acknowledge that even with vast amounts of information, the uncertainty of the future cannot necessarily be eliminated.

As a result, modern prediction no longer simply seeks to answer “what will definitely happen in the future,” but focuses more on the likelihood of different outcomes. For instance, medical predictions can evaluate the risk of a disease occurring, financial predictions can calculate the probability of an asset's price rising, and weather forecasts can provide the likelihood of rain. Predictions have shifted from a single answer to a description of various possibilities.

“It will rain tomorrow” and “there is an 80% chance of rain tomorrow” represent two different modes of prediction. The former delivers a judgment with certainty, while the latter explicitly conveys that other possibilities still exist. For someone planning a trip, what is truly valuable is often not an absolute answer, but knowing the likelihood of different outcomes to decide whether to bring an umbrella, adjust the itinerary, or prepare contingency plans.

In this sense, probability does not signify a lack of predictive ability, but rather an essential way in which modern prediction deals with uncertainty.

In other words, what modern prediction truly pursues is not the elimination of uncertainty, but rather the quantification of uncertainty—humans cannot make the future completely certain, but through data, models, and probability, they can provide a more precise description of the different outcomes that might occur in the future and their respective likelihoods.

3 From Prediction to Verification

When prediction results began to be expressed in terms of probabilities, a new question arose: if a model tells us that the probability of an event occurring is 80%, why should we believe this 80%?

Merely being able to provide a prediction result does not prove that a model is valid. Even if a model gives seemingly precise probabilities, these probabilities do not necessarily match the actual frequency of occurrence in the real world. Therefore, another crucial distinction between modern prediction and ancient empirical prediction is that prediction results must be put to the test of reality.

The most direct approach is to compare the model's prediction results with the actual outcomes that occur later. If the model assigns an approximately 80% probability of occurrence to a large number of similar events, then among the events corresponding to these predictions, the actual proportion of occurrences should also be close to 80%. If there is a persistent, significant discrepancy between the predicted probabilities and actual outcomes over the long term, it indicates that the model may have flaws and requires further analysis and calibration.

What needs to be noted here is that a single correct prediction does not prove that the model is valid; nor does a single incorrect prediction prove that the model is invalid. Probabilistic prediction inherently allows for errors to happen; what truly needs to be observed is the overall performance of the model across a vast number of predictions, and whether there is stable consistency between the predicted probabilities and the actual outcomes.

This fundamentally transformed the status of “rules.” In ancient empirical systems, an empirical rule passed down over a long time might gain authority simply because it was repeatedly used; but in modern science, experience alone is insufficient to serve as the ultimate basis. Whether a rule holds true still requires observation, experimentation, statistical analysis, and continuous predictive testing.

Going a step further, model validation also requires determining why it works, under what circumstances it works, and whether it remains valid once it moves away from the original data. If a model merely memorizes peculiar patterns in historical data but cannot adapt to new data, its high historical accuracy holds little significance. This is the common problem of overfitting in machine learning and statistical learning: the model may have overlearned incidental features in the training data without genuinely grasping stable rules that can generalize to new data.

Therefore, modern prediction is not a one-way chain of “observation → rule → prediction,” but rather a continuous loop of testing against reality and making revisions: observing data, discovering rules, building models, making predictions, then comparing prediction results with actual outcomes, and evaluating, revising, or even rebuilding the model based on the test results.

In this sense, the most critical difference between modern scientific prediction and ancient empirical prediction does not lie merely in using more data or more sophisticated computational tools, but in establishing a mechanism that subjects predictions to continuous testing against reality.

5 Machines Begin Learning How to Infer the Unknown

1 From Human Modeling to Machine Learning

Although the modern prediction models discussed in Chapter 4 are already capable of processing massive amounts of data and establishing relationships between variables through mathematical methods, in traditional modeling approaches, the fundamental structure of the model and the relationships among variables still generally needed to be designed in advance by humans. Humans needed to decide what problems to study, what data to collect, which variables to select, and how to describe the relationships between these variables. Machines were responsible for executing the calculations, while how the model should be constructed remained primarily human work.

This approach actually continued the tradition of human prediction discussed earlier. Whether it was the ancient practice of summarizing judgment rules based on experience, or modern science constructing mathematical models based on data, there was a shared link: humans were responsible for seeking noteworthy patterns in reality and organizing them into rules for explaining or predicting the unknown. The difference lies in the fact that modern science possesses more systematic data, mathematical tools, and validation methods.

However, as the scale of data continued to expand and relationships between variables became increasingly complex, humans began to encounter a new limitation: even with sufficient data, they might not necessarily be able to conceive of all the valuable relationships among variables in advance. Variables in reality may extend far beyond those humans can anticipate beforehand, and complex nonlinear relationships may also exist between different variables. Faced with such complexity, relying solely on humans to judge in advance “which variables are important” and “what relationships exist between them” could itself become a bottleneck for predictive capabilities.

Machine learning emerged precisely in this context. What it transformed was not merely computational speed, but the division of roles in the model-building process.

Traditional models generally follow:

现实问题
    ↓
人类选择变量
    ↓
人类设计模型
    ↓
机器进行计算
    ↓
得到预测结果

Machine learning, on the other hand, attempts to delegate part of the work of “finding patterns” and “building models” to machines:

现实问题
    ↓
人类准备训练数据
    ↓
机器学习数据中的模式
    ↓
形成预测模型
    ↓
输入新的数据
    ↓
产生预测结果

The most crucial change here is that machines no longer merely execute models designed by humans; instead, they begin to participate in building the models themselves.

Of course, this does not mean that machines can independently understand reality entirely detached from humans. Humans still need to define problems, prepare data, select training objectives, and design methods for evaluating models. What machine learning truly transformed was a part of the work in the model-building process: complex relationships that previously required humans to pre-design based on experience and theory can now be delegated to algorithms to a certain extent, allowing machines to learn automatically from large amounts of data.

For instance, when given a large number of historical medical cases and their corresponding diagnostic results, an algorithm can search for statistical relationships with the diagnostic outcomes from variables such as the patient's age, symptoms, test indicators, and their combinations—some of which may be patterns that humans could not have intuitively conceived of initially.

Consequently, an important shift occurred in how humans construct prediction models: in the past, humans discovered rules and then had machines compute them; machine learning, however, began to allow machines to discover rules from data, and then utilize these rules to make predictions.

2 From Machine Learning to Deep Learning

Machine learning solved an important problem: humans no longer need to write all the rules into the model in advance, but can let machines find patterns from data instead. Yet a limitation still remains—what kind of patterns machines are able to discover is still largely influenced by the way humans design features.

Traditional machine learning typically requires manual design or selection of features. For example, when processing data such as images or speech, humans often need to first extract features like color, edges, shapes, or acoustic spectra using specific methods, and then let the algorithm find patterns based on these features. For tasks where features are easy to extract and the problem structure is relatively clear, this approach is highly effective; however, real-world information often contains complex hierarchical relationships, and relying solely on handcrafted features makes it difficult to fully describe the information contained within.

The significant change brought about by deep learning is that it further reduced reliance on this manual feature engineering. Through multi-layered architectures, deep neural networks learn increasingly abstract representations layer by layer from raw data. For example, in image recognition, lower layers might learn edges and textures, intermediate layers gradually form shapes and local structures, while higher layers may form representations of object categories and their semantic features.

Therefore, deep learning actually pushed the aforementioned concept of “machines finding patterns” a step further:

传统机器学习

原始数据
    ↓
人工设计或选择特征
    ↓
机器学习特征之间的关系
    ↓
预测结果


深度学习

原始数据
    ↓
机器逐层学习特征表示
    ↓
形成多层次抽象表示
    ↓
学习复杂模式
    ↓
预测结果

What is truly important here is not “depth” itself, but that the model begins to simultaneously learn “what should be paid attention to” and “what relationships exist between these pieces of information.” In the past, humans needed to first convert reality into features suitable for machine processing before letting the machine learn; deep learning, however, begins to let the machine participate in this conversion process itself.

While this enables the model to process more complex data, it also introduces a new issue: the internal representations formed within the model may become increasingly difficult to explain using concepts familiar to humans. The fact that a model can provide accurate predictions does not mean that humans can clearly explain precisely which internal features it relied upon to arrive at that result.

As a result, predictive capability and human comprehension began to decouple to a certain degree. In the past, across many traditional prediction models, the variables and relationships utilized by the model typically had relatively clear definitions, allowing humans to broadly understand how prediction results were generated through these variables and model structures. In deep learning, however, models can form massive, hierarchically complex internal representations from data, many of whose features and relationships do not directly correspond to concepts that humans can name.

This signifies that the way humans construct prediction models has taken another step forward: machines do not merely compute for humans, nor do they merely look for rules for humans; they have begun to form complex internal representations.

And when this capability further expands to different types of information such as text, images, and audio, the question evolves accordingly: can machines establish more generalized representations and forge connections across different types of information, thereby handling even more complex tasks?

3 From Specialized Models to General-Purpose Models

Many early deep learning models were primarily used for specific tasks such as image recognition, speech recognition, and text classification. These models were typically trained on well-defined problems with relatively fixed inputs and outputs; thus, the capabilities of the models were often tightly bound to specific tasks.

As training data became increasingly abundant and model scales continued to expand, researchers gradually discovered that the capabilities learned by models from data did not necessarily have to be confined solely to the tasks for which they were initially intended. When exposed to sufficiently diverse information, a model might learn features and information representations that can be shared across different tasks, thereby harboring the potential to tackle a variety of problems.

Large language models are a typical example of this shift. Their most fundamental training objective appears remarkably simple: predicting the next token based on the preceding text. A token is not entirely equivalent to a word in everyday language, but rather the basic unit used by the model when processing text. However, when the model continuously makes such predictions across vast and diverse text data, in order to improve prediction accuracy, it cannot merely memorize simple correspondences between words; it must also gradually learn language structures, contextual relationships, and the diverse knowledge embedded within the text.

Therefore, although the training objective remains predicting the next token, the capabilities the model ultimately acquires are no longer confined to this single task. Once trained, the same model can be further applied to various tasks such as question answering, translation, summarization, and code generation.

This means that the development of models has begun to shift from “building a model for a specific problem” toward “addressing multiple problems with a single general-purpose model.”

Traditional dedicated models can be simply represented as:

特定训练数据
      ↓
针对特定任务训练
      ↓
   专用模型
      ↓
   特定任务
      ↓
   预测结果

Whereas general-purpose models are closer to:

海量、多样化训练数据
        ↓
    通用模型训练
        ↓
      通用模型
        ↓
  ┌─────┼─────┐
  ↓     ↓     ↓
问答   翻译   代码

Note: "General-purpose" here does not mean that the model can solve all problems, but rather that the same model can adapt to multiple different types of tasks given different inputs and task requirements.

The most critical shift here is not that the model suddenly acquired a specific new feature, but that the capabilities learned by the model began to exhibit cross-task transferability. Instead of requiring a complete, separate model for every single problem, different types of problems can now be addressed on the foundation of an already-trained general-purpose model.

This also gives “prediction” a new role throughout the model-building process: it is no longer just the final output result of a dedicated model, but can also serve as the foundation for training general-purpose models and forming generalized capabilities.

6 When Machines Begin Simulating the Future

If machines are already capable of learning patterns from massive amounts of information and forming general-purpose models that can tackle diverse problems, we might take this a step further with a bold conjecture: as this capability continues to evolve, will machines be able to build dynamic models of the real world and use them to simulate different states that may arise in the future?

A relatively intuitive direction is the personal digital twin.

Today, we can already construct a digital profile of an individual through their search history, consumption habits, location changes, social relationships, and other digital traces. However, while a digital profile primarily describes “what this person is like,” a digital twin seeks to take it further by establishing a dynamic model that continuously mirrors the real-world subject: when the state of the real-world subject changes, the digital model updates accordingly and simulates the changes that might occur under various conditions.

Extending this concept to individuals, a system could continuously gather information such as a person's behavioral patterns, shifting interests, financial situation, social connections, and past decisions to build a constantly updated personal model, attempting to simulate the actions this person might take under different conditions. At this point, what the machine predicts is no longer just an isolated outcome, but potentially a trajectory of future development.

This kind of dynamic simulation can be further extended to enterprises, cities, and even broader social systems. Compared with personal models, these systems not only involve far more variables, but also encompass complex interactions between different agents. As models gradually gain the ability to describe these relationships and dynamic shifts, what machines do may also evolve from predicting a single outcome to simulating futures that could unfold under different conditions based on the current state.

Of course, this remains a forward-looking vision. Real-world individuals and societies contain numerous unobservable variables, and human behavior is subject to chance occurrences and the actions of others. Therefore, even if future AI can build highly intricate personal or societal models, it will still be difficult to arrive at a truly definitive future. Nevertheless, being unable to determine the future does not mean that none of its possibilities can be deduced in advance. While machines might not accurately determine whether a specific event will definitely happen, they could potentially project certain plausible developmental paths in advance by simulating variations under different conditions.

Science fiction has already offered extreme visions of this possibility. In the Tom Cruise film *Minority Report*, humanity invented intelligent beings known as “Precogs” capable of detecting brainwaves and predicting criminal intent. They could detect crimes an individual might commit in the future ahead of time, allowing the police to take action before the crime actually occurred. Although the predictive capabilities depicted in the film far exceed those of real-world AI, it raises a question well worth contemplating: if machines can build sufficiently complex models of human behavior, could certain future states be deduced before they ever happen?

As this capability continues to advance, “predicting the future” might also transform: it will no longer merely be about offering a probability based on historical data, but could become a continuous simulation within machine-built models of the pathways reality might follow, and where different choices might ultimately lead reality.