Looking back at the history of artificial intelligence, it's not hard to see that despite the astonishing progress in performance over recent decades, we still seem to be caught in a fundamental "trap." This trap concerns the true definition of intelligence, and to understand it, we must ask: Where does AI's knowledge truly come from?
Phase 1: "Force-Fed" Rules and Heuristics
In the beginning, scientists tried to use algorithms to "brute-force" searches for answers. In simple games, for example, a computer could exhaust every possible move to find the optimal solution. At this stage, "intelligence" was nothing more than computational speed; it involved no knowledge.
When the scale of the problem expanded, such as in chess or Go, "brute force" failed. Thus, more sophisticated algorithms like A* search were born. These algorithms required human experts to write an "evaluation function" (a Heuristic Function) to guide the search, telling the computer which move "looked" better.
Note that the "intelligence" here was still a direct translation of human wisdom. This heuristic function was the concentration of human experience, a distillation of human "knowledge" into "rules," which were then forcibly "stuffed" into the system.
Phase 2: "Copied" Expert Experience
Next came the wave represented by "expert systems." People invited experts from various fields—like doctors and engineers—to write down their lifetimes of accumulated judgment and solutions as a series of "if...else..." rules. These rules were entered into a "Knowledge Base."
In this way, a computer could behave like an "expert" in a specific domain. But its essence was still a replica of human knowledge. The computer was merely an efficient "inference engine," executing logical deductions within a knowledge framework preset by humans. The source of the knowledge was, 100 percent, human.
Phase 3: Human Knowledge Distilled from "Big Data"
It wasn't until the advent of neural networks and deep learning that people felt, for the first time, that machines might finally be "learning on their own." They no longer relied on rigid rules but instead "trained" on massive amounts of data to acquire abilities, such as "learning" what a cat is after seeing thousands of cat photos.
But we must think more deeply: Where does this data come from?
In "Supervised Learning," this data consists of photos taken by humans, text written by humans, and sounds recorded by humans. And most critically, it has all been "labeled" by humans (e.g., "This is a cat").
These "problem-correct answer" data pairs inherently contain vast, unrefined human knowledge. What the machine does is "distill" patterns and correlations from this data, which is saturated with collective human wisdom. Humans are no longer the direct authors of the rules, but rather the "curators" and "labelers" of the knowledge. The source of the knowledge is still humanity itself.
Phase 4: The More Hidden "Human Knowledge" of Unsupervised Learning
Unsupervised learning seems to go a step further, as it no longer requires humans to painstakingly "label" data. But its essence has not changed at all.
The material it "learns" from—such as Wikipedia, the entire internet's text, countless books—what is this? It is still the crystallization of knowledge that humans created for other purposes (like communication, record-keeping, and inheritance).
The cleverness of unsupervised (or self-supervised) learning lies in its method of "manipulating" this human data to automatically "create problems." For instance, in language models, it masks a word in a sentence ("The sky is ___ color") and then has the model guess the word. To guess correctly, the model is forced to learn human grammar, logic, and even common sense.
This is merely a different way of "feeding" human-created knowledge to the machine. The root is still the knowledge accumulated by human civilization.
The Closed Loop: Imitating Knowledge vs. Creating Knowledge
This forms a perfect closed loop: What humans teach machines is, invariably, knowledge that humans already possess.
In the "learning" process, the machine does not truly "understand" the content it processes. It is merely finding an efficient statistical mapping relationship within the confines of human-defined datasets and tasks. It is more like a very well-read student who has memorized every book but cannot "think" about a completely new problem not mentioned in those books.
True intelligence should possess the ability to autonomously learn new knowledge from the environment.
A human infant does not need massive "labeled data" to be told what "gravity" is. He will explore, observe, and test on his own (such as by throwing a toy on the floor). In this cycle of "interaction-error-summary" with the real world, he gradually comes to understand the physical laws of this world. This is the "creation" of knowledge, from 0 to 1.
Today's artificial intelligence is still far from being able to do this.
Perhaps in the future, when we can make machines truly "Embodied" in the real world—allowing them to actively explore, experience failure, and generalize from experience—only then might they truly "learn to learn."
That day, we might have reason to say that it is not "imitating" knowledge, but "creating" knowledge. And that day, perhaps, will be the beginning of our escape from the "illusion of intelligence" and the start of our journey toward true intelligence.