A model sits at the foundation of how we know anything at all. Every model is an abstraction of some object, process, or history of change. In physical terms, that means at least four dimensions are always involved: three dimensions of space, plus time. But for something to count as a thing in the first place, it must be distinguishable from other things—or from empty space. That already implies describability and a degree of stability. A thing cannot be a duck today and corn tomorrow. It must persist enough in space and time to be repeatedly identified.
Even that seemingly basic idea depends on a semantic model. Concepts are built through language, and different languages emphasize different aspects of the world. When concepts are vague or misaligned, people can argue past each other indefinitely, like one talking about chickens and the other about ducks. Much of the resistance to metaphysical discussion comes from this: some questions do not have a single agreed definition to begin with.
What does it mean for something to exist stably? At minimum, it means the description can refer to it again and again. The thing may be static, periodic, randomized, or complex. The last two are not the same. A randomized thing is often difficult to describe in detail but can still be described statistically as a whole. A complex thing, by contrast, contains structure, regularity, or patterns. All four kinds can be generated in a discrete framework such as cellular automata, but the actual world is far more complicated than that.
The regularities of things themselves are the subject of physics, from subatomic particles to the entire universe. Because physical laws exist, the world becomes knowable and, to some extent, predictable. The part of physics concerned with molecular transformation became chemistry; the part of chemistry concerned with molecules involved in life became biology. As one moves from physics to biology, the scope of the laws narrows while their descriptive precision increases. The more universal a law is, the more it can resemble a useful platitude. Electromagnetic interaction is fundamental, but when describing protein folding, a model based on hydrogen bonds is usually easier to understand and more practical.
As disciplines become more specialized, they build new concepts around the properties of things in the physical world. Ideally, those concepts are tools for communication and description, not barriers between fields. In fact, a physical model can be seen as a restricted form of mathematical logic, or mathematical logic with boundary conditions attached. Mathematics allows negative numbers without end; physics does not let you write down minus one thousand degrees Celsius and call it sensible. Logic is clean and orderly, which makes it useful for organizing thought, but reality sometimes presents facts that seem completely illogical. When that happens, facts do not need to be corrected—our logic probably omitted something important.
In that sense, causal analysis is a powerful tool for studying mechanisms, but if the data are biased or the relationships are deeply entangled, then excessive simplification becomes dangerous. Many arguments look rigorous and well structured while the facts themselves may unfold along an entirely different path. No model should be allowed to drift free from the world it is meant to describe.
At root, every model of the real world is grounded in a physical model. Such a model may describe a single object, or the relations among two or more objects. From that basis, one can build thermodynamic models that describe stable states, or dynamic models that describe change across time scales. Yet the history of physical science has left it with a strong preference for bottom-up analysis: break things apart, describe the parts, then recombine them. That approach often fails to capture reality as lived or observed. Systems thinking, which is more synthetic and heuristic, tends to begin from relationships and interactions. And in many cases we do observe similar patterns across very different domains—power-law distributions are a familiar example.
Whatever form it takes, a model is a simplified path of thought about the world. If one starts forcing reality to fit a model merely for the sake of having a model, then the whole exercise has been inverted: something simple is made unnecessarily complicated. At the same time, simplification is itself a subjective matter. Fortunately, modern humans share a common evolutionary background, so many of our subjective judgments overlap enough to gain a kind of objective standing. Under a different intelligent civilization, things might look very different.
Why do we use models at all? Most likely because they reduce the cost of thinking. They extract commonality and regularity from the overwhelming variety of reality, and that is crucial for survival. Without models, we would need extraordinary memory just to record and revisit our personal histories in full detail. Imagine a tree covered in fruit. If the brain could not directly identify it as an apple tree, we would have to memorize the shape of each branch and the appearance of every fruit hanging from it, even if all the apples looked broadly alike. The concept of an apple tree is a simple cognitive model. It rescues us from noise and detail.
That also means every model abstracts concepts and discards information. Knowing there is an apple tree ahead does not tell you which leaf has an insect on it. That information exists in the object itself, but the evolved brain quietly throws it away. The relationship among the real world, models, and human cognition is layered:
$$真实世界 \= 认知模型+模型未捕获信息 \= 人的认知+人对认知模型的理解偏差+模型未捕获信息$$
At the very least, human beings know through a physiological model built from the nervous system. More specifically, we know through language models. If language describes poorly, distortion follows. Combined with the abstraction inherent in the model itself, what people can know about the world always includes both the model's bias and their own bias. Personal bias can be reduced through learning or by introducing mathematical description. Model bias is much harder to reduce.
A simple example is regional prejudice. It functions as a crude cognitive model: ask where someone is from, then jump directly to a judgment about them. It may fail to reflect reality, yet it remains common. Popularity does not make it reasonable.
Models also differ in robustness. A simple model is often robust because it is insensitive to fluctuations in the facts, but that same insensitivity makes its description coarse and one-dimensional. A fine-grained model captures subtle changes in reality, but because it is so sensitive, its output may swing dramatically. Statistics has a similar idea: the bias-variance tradeoff. Optimizing a single model means finding a balance between overfitting and underfitting reality, then extracting useful parameters. The same problem appears in human understanding. Regional prejudice is a kind of underfitting; empiricism, in many cases, tends toward overfitting.
So how can model bias be reduced? The simplest answer is to master multiple independent models. The biases and blind spots of any one model can be offset or covered by the independent parts of others, producing a fuller and more three-dimensional understanding.
In historical research, trying to explain events with one single "-ism" usually leads to pseudo-patterns. It is like carrying a hammer and then treating every usable fact as a nail, while historical facts that are more like nuts and require a wrench are simply ignored. Using several historical perspectives at once is often more helpful in reconstructing what actually happened. Every model departs from the facts in some way. Simpler models are easier to accept because they align with Occam's razor, but in the larger view, an ideal model would be built by weighting and integrating several independent models. That is the point at which a model most closely reflects reality.
Of course, people do not always want reality itself; sometimes what they actually want is the result after reality has been filtered by a model. And truly independent models may be rarer than they seem. Many are correlated from the start, or share a deeper underlying consistency. Integrating such models may reveal little that is genuinely new. If two models are sequentially connected, the effect of their combination should resemble the product of independent model effects rather than their sum. If two models describe independently, then their joint effect is more like a weighted combination of single-model performance, with the weights determined by the problem at hand.
$$完美模型 \= 独立模型1 + 独立模型2 + … +独立模型n$$
A model can only approach perfection; it can never become perfect. There will always be unknown information or undiscovered perspectives. Still, current developments in artificial intelligence may eventually produce models whose descriptive power exceeds the level of human understanding. We may never fully grasp how such a model works, and yet its predictive ability could be astonishingly good. From an engineering standpoint, that may not be a problem. Ethically, however, modern humans will probably still want their feelings considered.
Adding more models can move the system's optimum toward parts of reality that were once dismissed as noise. More facts and more data can have a similar effect. At the moment, there still seems to be too little attention given to integrating more models.
This, then, is itself a cognitive model of models. For an individual, broader experience and wider horizons help reveal the world's complexity and accumulate more usable experience. Openness to multiple perspectives makes it easier to discover integrated models that fit the facts more closely. Not every independent model deserves inclusion—many may be irrelevant to the question at hand, and some may only add variance. In plenty of situations, simple rules are the most practical and effective tools. But during the exploratory stage of cognition, using multiple models can help produce an understanding that exceeds the limits of any single person or perspective.