From a Thought Experiment to a New World

Have you ever wondered where the AI we use every day, such as voice assistants on our phones, recommendation algorithms, and even the recently popular ChatGPT, comes from? Many people might think that AI was "invented" by some genius in a lab. But the truth is far more exciting. AI is not an isolated invention; it is more like a "relay race of ideas" that has spanned nearly a century. Its starting point is not even a line of code, but a profound philosophical question: "Can machines think?" This question is like a stone thrown into a calm lake, creating ripples among generations of brilliant minds. Since then, philosophers, mathematicians, engineers, psychologists, and countless pioneers have dedicated themselves to this field. Some defined its name, some paved its path, some persevered during the winters, and others ignited its energy today. In this article, we will take a different perspective and connect the key turning points of AI from 0 to 1 through 10 of the most representative figures. You will see:

• How a great dream was "named" and "defined."
• How the two technical routes of "symbolism" and "connectionism" clashed and merged.
• How the three "fathers of deep learning" persevered during the winter and ultimately welcomed the revival of the entire field.

Let "intelligence" move from philosophy to science.

Any great technological revolution often originates not from a specific invention, but from a groundbreaking question. The starting point of artificial intelligence (AI) is especially so. Its story does not begin with a roaring machine or a line of magical code, but from a thought experiment proposed on paper by a genius mathematician. This person and his question brought the concept of "intelligence," which had lingered in the halls of philosophy for thousands of years, into the arena of modern science for the first time. He is Alan Turing.

In 1950, the dawn of computer science had just arrived, and the machines of the time were bulky and slow, capable only of performing the most basic calculations. However, Turing's thoughts had already transcended the limitations of the era. In his groundbreaking paper "Computing Machinery and Intelligence," he posed a seemingly simple yet profoundly deep question: "Can machines think?" Turing keenly realized that directly discussing the definition of "thinking" would lead to an endless philosophical quagmire. Therefore, he cleverly transformed it into a verifiable game—the "Imitation Game," which later became world-famous as the "Turing Test."

The rules of the game are as follows: a questioner communicates through text with two anonymous entities, one of which is a human and the other a machine. If, after a sufficient amount of time, the questioner cannot distinguish which is the machine, then we can say that the machine has passed the test and exhibited intelligent behavior indistinguishable from that of a human. This is what makes the starting point of AI so unique: it is not an "invention" aimed at solving specific tasks, but a "challenge" aimed at answering fundamental questions.

The greatness of the Turing Test lies in its provision of an operational and evaluative standard for the vague concept of "intelligence." It no longer gets bogged down in whether machines possess souls or consciousness; instead, it focuses on their external behavioral performance. It's like saying we don't need to open a black box to explore its internal structure; we can judge its capabilities simply by observing its output. This pragmatic thought transformed a purely philosophical speculation into an engineering goal that engineers and scientists could tackle.

Turing planted a seed of thought, but to let it take root and sprout, it needed fertile soil and a passionate gardener. This person soon appeared: John McCarthy. In 1955, Turing had already passed away, but his question inspired a group of young scholars across the Atlantic. At that time, research on "thinking machines" was scattered across various fields, with names like "Cybernetics," "Automata Theory," and so on.

Young Dartmouth College mathematics assistant professor McCarthy felt that these scattered sparks needed to be gathered into a blazing flame. He planned to organize a several-week-long seminar in the summer of 1956, inviting the brightest minds in the United States to explore the possibility of simulating human intelligence with machines. In the conference proposal drafted with Marvin Minsky and others, they expressed a genius-like optimism, declaring that "every aspect of intelligence can, in principle, be precisely described, allowing machines to simulate it."

To give this new field a clear identity, McCarthy needed a catchy name. He deliberately avoided the influential term "Cybernetics" at the time because he did not want this new field to be seen as part of the academic territory of Norbert Wiener, the founder of Cybernetics. McCarthy later recalled that he created this new term to draw a line and establish an independent academic identity. The name he carefully selected was—"Artificial Intelligence."

The Dartmouth Conference of 1956 thus became the "Genesis" of AI history. It not only officially named the discipline but, more importantly, it brought together a group of thinkers with a shared dream and established the initial research agenda. At the conference, Allen Newell and Herbert Simon demonstrated the world's first "thinking" program—the "Logic Theorist," which could prove mathematical theorems using symbolic logic like a human, greatly inspiring the attendees.

The birth of the name "Artificial Intelligence" meant that a new continent had been officially discovered. It provided all explorers passionate about "machine intelligence" with a common identity and a unified banner. From then on, they were no longer isolated mathematicians, psychologists, or engineers, but "artificial intelligence scientists." McCarthy not only named the discipline but also created the Lisp language in 1958, a powerful symbolic processing tool that became the "official language" of early AI researchers, allowing them to truly transform abstract logic and ideas into programs that could run on machines.

From Turing's philosophical "question" to McCarthy's disciplinary "name," artificial intelligence completed a crucial leap from 0 to 1. Turing defined the ultimate goal, while McCarthy sounded the rallying call, initiating one of the grandest and most exciting scientific journeys in human history. This journey began with the philosophical inquiry of "who we are" and ultimately led to the scientific practice of "creating new intelligence" through code and algorithms.

The First Clash of Two Roads: Genius Optimism vs. Reality's Cold Water

Why were early AI scientists so optimistic?

In the dawn of artificial intelligence, the entire field was permeated with an almost fanatical optimism. The core of this confidence stemmed from a simple yet powerful belief—symbolism. Led by Marvin Minsky of MIT, the first generation of AI scientists firmly believed that human intelligence, and indeed all intelligent activities, could be broken down into a series of logical symbols and formal rules. In their view, the brain was merely a "meat machine," and as long as we could find the right rules, we could reproduce the process of thinking on a computer.

This belief was not unfounded; it was built on a series of exciting early successes. In 1956, a program called "Logic Theorist" emerged, widely regarded as the first true artificial intelligence program. This program not only successfully proved 38 theorems from the famous mathematical work "Principia Mathematica" but even found more elegant proofs for some of them than the original text. One of its creators, Herbert A. Simon, excitedly proclaimed, "We have invented a computer program that can perform non-numeric thinking, thus solving the ancient mind-body problem." This achievement demonstrated to the world that machines could indeed accomplish creative intellectual tasks once thought to be exclusive to humans. Another famous example is the SHRDLU system, which could understand and execute complex tasks in a virtual block world through natural language commands, such as "Put that red pyramid on the blue block."

These successes in the "toy world," though limited in scale, illuminated the path to general intelligence like a beacon. It was these tangible results that greatly encouraged Simon, Minsky, and others. They made predictions that seem extremely bold today; for example, Simon once predicted that within ten years, machines would be able to defeat human world chess champions and discover new important mathematical theorems. During that "golden age," people generally believed that as long as they continued down the path of symbolism, achieving machine intelligence on par with humans was merely a matter of time.

How did the first AI winter come about?

However, the optimism of the geniuses soon collided with a cold wall named "reality." When AI researchers attempted to apply those programs that worked well in the "toy world" to the real, complex world, the fundamental problems of symbolism were laid bare. First, symbolism struggled to handle the ubiquitous "common sense" and "uncertainty" present in the real world. Human daily decision-making is filled with ambiguity, intuition, and default background knowledge, all of which are extremely difficult to encode into precise logical rules.

For instance, we all know that "water is wet" and "birds can fly," but manually inputting these endless common sense facts into a machine is nearly an impossible task. Secondly, AI systems face a fatal obstacle when scaling up—"combinatorial explosion." This means that when the variables of a problem increase even slightly, the number of possibilities the system needs to compute can grow exponentially, quickly exceeding the processing power of any computer.

For example, the SHRDLU system's performance would dramatically decline and become impractical once its "block world" became slightly more complex. These fundamental limitations caused AI development to lag far behind the initial promises. Disappointment began to spread, culminating in 1973 with a document known as the "Lighthill Report." This report, commissioned by the British government, sharply criticized, "To date, no results in any field have achieved the significant impact that was originally promised."

The report pointed directly to the failures of AI research in solving real-world problems, particularly its inability to address the "combinatorial explosion" issue, concluding that much of the foundational AI research was not worth continuing to fund. The publication of this report directly led to significant cuts in funding for AI research by the British government, forcing many university AI labs to close. This chill also affected the United States across the ocean, where funding agencies became cautious and leaned towards short-term projects with clear application prospects.

Thus, due to the vast gap between promises and reality, artificial intelligence faced its first "winter." Even Minsky later admitted that their "biggest mistake... was not realizing how difficult the problems we were trying to solve were."

Why is teaching AI to "accept uncertainty" considered a major breakthrough?

As the path of symbolism hit a dead end, another entirely different line of thought brought new hope to AI. The pioneer of this new path was Turing Award winner Judea Pearl. He led a "probability revolution," whose core idea was: rather than forcing AI to understand the world with black-and-white logic, it is better to teach it how to accept and process "uncertainty." Pearl's revolutionary weapon was the "Bayesian networks" he proposed in the late 1980s.

This is a clever graphical model that can represent the probabilistic dependencies between different variables using intuitive graphical structures. More importantly, it provides a rigorous mathematical method that allows AI to dynamically update its "belief" about the likelihood of events based on new evidence. This exhibited tremendous power in fields like medical diagnosis. Traditional expert systems attempted to diagnose using rigid "if... then..." rules, such as "If the patient has a fever, then they may have the flu."

But reality is far more complex: a fever could also be a symptom of other conditions, and the strength of the association between each symptom and disease varies. Such systems based on absolute rules often become very fragile when faced with incomplete or contradictory information. In contrast, the Bayesian network approach is entirely different. It can construct a network of probabilistic relationships that includes multiple diseases and symptoms. When a doctor inputs the evidence "the patient has a fever," the system does not arrive at an absolute conclusion but automatically updates the probabilities of all related diseases (like flu, pneumonia, etc.) based on Bayes' theorem.

If new evidence "the patient is coughing" is inputted, the system recalculates and further adjusts the probability distribution, providing a more realistic, probability-based diagnostic suggestion. This shift from pursuing "certainty" to embracing "uncertainty" represents a significant intellectual advancement. It allowed AI for the first time to engage in reasonable reasoning and decision-making in the real world, which is often filled with incomplete information and ambiguity. Pearl's work not only provided powerful new tools for AI to escape the dilemmas of reality, widely applied in fields like healthcare, speech recognition, and fault diagnosis, but more importantly, it opened a new path for the development of artificial intelligence toward more powerful intelligence.

Persevering through the winter: The Revival of Neural Networks and the "Three Giants"

When the optimistic wave of symbolism receded and AI research entered a long and cold "winter," most researchers and funding shifted to more practical fields like expert systems. However, on the fringes of academia, a small group of people remained convinced that the path of connectionism and neural networks, which had been criticized by Marvin Minsky and nearly abandoned, was the correct direction toward true intelligence. They were the steadfast guardians in the winter, a minority within a minority.

It was this almost stubborn belief that ultimately ignited the second revolution in artificial intelligence. The leaders of this group were later hailed as the "three giants of deep learning": Geoffrey Hinton, Yann LeCun, and Yoshua Bengio.

What exactly is "deep learning"?

To understand the contributions of these three scientists, we first need to answer a fundamental question: what exactly is "deep learning"? How does it fundamentally differ from early neural networks? Early neural networks, such as perceptrons, had very simple structures, usually consisting of only one or two layers. This is like a child just learning to draw, who can only recognize very basic lines and color blocks. If you want them to recognize a cat, you must first manually tell them the features of a cat—"pointy ears," "whiskers," "round face."

This process is called "feature engineering," which is time-consuming and labor-intensive, and often yields poor results because the real world is far more complex. Deep learning, as the name suggests, is centered on "depth"—it uses neural networks with many layers (from a few to hundreds). This multi-layer structure gives it a powerful ability: to automatically learn features. We can use a more vivid analogy to understand this: it is no longer teaching a child to draw but providing them with a complete visual cortex system.

When seeing a picture of a cat, the first layer of this "deep" network might automatically learn to recognize the most basic edges and corners; the second layer, based on the results of the first layer, learns to combine them into more complex shapes like eyes and ears; and deeper layers learn to recognize the concept of a "cat face" or even an entire "cat." The entire process is end-to-end, from raw pixel points to the final conclusion of "cat," with the machine learning autonomously throughout, without humans needing to define what "pointy ears" or "whiskers" are.

This learning method, progressing from the specific to the abstract, layer by layer, is the most essential difference between deep learning and early neural networks, and it is the source of its power.

The "Three Giants": The Minority that Ignited the Flame in the Winter

It was the shared belief in this "depth" that tightly connected Hinton, LeCun, and Bengio. During the decades when neural networks were neglected, they faced skepticism from the academic community, like three lonely torchbearers, each fighting in different directions yet resonating with each other, ultimately solving the core problems that allowed deep learning to transition from theory to reality.

• Geoffrey Hinton: The Founder of Making Deep Networks "Trainable" Hinton is known as the "father of deep learning," and his greatest contribution was solving the fundamental problem of "how to effectively train a deep network."

In 1986, he and his collaborators popularized the backpropagation algorithm. This algorithm acts like a strict teacher; when the network makes an incorrect judgment, it can "backpropagate" the error signal from the last layer back through each layer, telling each neuron's parameters how to adjust slightly to do better next time. This breakthrough made training multi-layer neural networks possible and laid the foundation for the entire deep learning field.

• Yann LeCun: The Pioneer Who Made AI "See" the World LeCun focused on how to enable machines to "see" the world. He realized that processing images could not be treated the same way as processing ordinary data. Inspired by the biological visual cortex, he developed convolutional neural networks (CNNs) in the late 1980s.

CNNs mimic the way the eyes capture local information through "convolutional kernels" and significantly reduce the number of model parameters through "weight sharing," making them both efficient and precise when processing images. The LeNet-5 network he designed in 1998 was successfully applied to a bank's check handwriting recognition system, becoming a model example of CNN's first commercial application and paving the way for subsequent breakthroughs in computer vision.

• Yoshua Bengio: The Theorist Who Made AI Understand "Language" While LeCun taught AI how to "see," Bengio pondered how to make AI "read" and "understand."

He focused on solving the "curse of dimensionality" problem in natural language processing (NLP). His proposed neural probabilistic language model innovatively introduced the concept of word embeddings. This technology maps each word to a high-dimensional continuous vector space, ensuring that semantically similar words are also close in space. For example, the vectors for "king" and "queen" would be very close. This allowed machines to capture semantic relationships between words for the first time, laying a solid foundation for the development of all sequence models, such as machine translation and sentiment analysis.

These three scientists each solved "how to learn" (backpropagation), "how to see" (CNN), and "how to understand" (word embeddings), and their work complemented each other to build the core technological landscape of modern deep learning.

2012 ImageNet: The "Big Bang" that Ignited the Revolution

Although the "Three Giants" had already prepared the theoretical gunpowder, igniting this revolution required a decisive moment. That moment arrived in 2012. The ImageNet Large Scale Visual Recognition Challenge (ILSVRC) was the "Olympics" of the computer vision field at the time, requiring participating algorithms to recognize and classify over a million images across 1,000 categories. Before 2012, the champions of the competition had always been teams using traditional machine learning methods, and improvements were always hard-won.

However, everything changed that year. Hinton and his two students—Alex Krizhevsky and Ilya Sutskever—entered the competition with a deep convolutional neural network called AlexNet. AlexNet is an 8-layer deep network that not only adopted LeCun's CNN architecture but also creatively used new techniques like ReLU activation functions and Dropout to enhance performance and prevent overfitting, leveraging the powerful computing capabilities of two GPUs for parallel training.

The results were groundbreaking. AlexNet won the championship with a Top-5 error rate of 15.3%, while the second-place score was 26.1%. This more than 10 percentage point gap caused a seismic shock throughout the AI community. It was no longer a minor improvement but a dimensional crushing. This victory irrefutably proved that with sufficient data and computational power, the performance of deep learning far exceeded all traditional methods. The 2012 ImageNet competition is thus recognized as a watershed event in AI history, marking the "ignition point" of the deep learning revolution.

As Hinton said, from then on, "almost all computer vision research turned to neural networks." This victory was like a starting gun, announcing the complete end of the AI winter and the beginning of a new era dominated by deep learning. Those who had persevered in the darkness for decades finally welcomed their dawn.

From the Laboratory to the World

The Creation, Popularization, and Reflection of AI If the three giants of deep learning found a powerful engine for artificial intelligence, the story after the 2010s is about how to connect the steering wheel to this engine, lay down the roads, and ultimately reflect on where it will take humanity.

This process was also driven by several key figures, each answering three core questions: How does AI learn to "create"? How does AI step out of the ivory tower? And when AI possesses immense power, how do we harness it?

The Qualitative Change of "Generative AI": Ian Goodfellow and the Inspiration from a Bar Debate In 2014, Ian Goodfellow, then a PhD student in Montreal, gathered with friends at a bar.

An academic debate about how to make computers generate realistic photos sparked his inspiration. That night, he conceived a genius framework—Generative Adversarial Networks (GANs). The essence of this idea lies in "adversarial." The GAN system consists of two neural networks that play against each other: a "generator" and a "discriminator."

The generator's task is to act like a skilled "forger," continuously learning the features of real data and then creating "counterfeits" (like a fake human face photo) that are convincing enough to deceive. The discriminator plays the role of an "authenticator," with the sole goal of accurately distinguishing which data is real and which is fabricated by the generator. The training process is like an endless zero-sum game: the generator strives to deceive the discriminator, while the discriminator works hard to uncover the deception.

In this escalating confrontation, both evolve together. Ultimately, when the discriminator can no longer effectively distinguish between real and fake, it means the generator has mastered the ability to create highly realistic data. This idea was so novel and powerful that Yann LeCun, one of the three giants of deep learning, praised it as "the most interesting idea in machine learning in the past 20 years." The birth of GAN marked a qualitative change in the history of AI development.

Before this, AI was more like a diligent "recognizer," adept at classification, recognition, and prediction. GAN, however, endowed AI with the identity of a "creator," giving it the ability to generate new, complex content from scratch for the first time, directly opening the door to the era of AIGC (Generative Artificial Intelligence) that we know today.

The Key to Technology Popularization: Andrew Ng as the Evangelist

A revolutionary technology, if it remains confined to the laboratory, ultimately has limited value. Andrew Ng played a crucial "evangelist" role in transforming AI from a tool for a select few elites into a skill that millions around the world can learn and apply. As a professor at Stanford University and co-founder of Coursera, Ng launched the online course "Machine Learning" in 2011, which became the catalyst for the MOOC (Massive Open Online Course) wave, attracting millions of registrations worldwide.

Subsequently, the "Deep Learning Specialization" course he launched in collaboration with DeepLearning.AI and the "AI for Everyone" course aimed at non-technical individuals further lowered the barriers to acquiring AI knowledge. By 2023, over 8 million people had participated in his courses. Ng not only popularized knowledge but also proposed a far-reaching idea: "AI is the new electricity."

He believed that just as electricity revolutionized agriculture, transportation, manufacturing, and nearly every industry a century ago, AI is now reshaping various sectors with unprecedented power as a foundational technology. This metaphor accurately points to the future of AI—it is not an isolated product but an infrastructure that will permeate all aspects of society. It is this foresight regarding the engineering and industrial application of AI that has greatly accelerated the transition of AI from academic research to industrial practice.

The Alarm of AI Ethics

Timnit Gebru and the Inescapable Reflection As the power of AI grows exponentially and begins to deeply intervene in social decision-making, a serious question arises for everyone: How can we ensure that this technology is fair, just, and responsible? AI ethics scientist Timnit Gebru has become one of the most representative "whistleblowers" in this field. In 2018, Gebru and her collaborators published a landmark study titled "Gender Shades."

They found that mainstream commercial facial recognition systems at the time exhibited severe biases: the accuracy rate was nearly perfect when identifying lighter-skinned males, but the error rate soared to nearly 35% when identifying darker-skinned females. This study served as a wake-up call, powerfully revealing how biases in training data could be amplified by AI systems, leading to systemic injustices against marginalized groups. This research directly prompted companies like IBM and Microsoft to improve their algorithms to reduce bias.

A few years later, while serving as co-lead of Google's ethical AI team, Gebru found herself in the spotlight again due to a paper titled "The Dangers of Random Parrots: Can Language Models Be Too Big?" This paper sharply pointed out the biases, environmental costs, and risks of large language models, which only mimic human language patterns without truly understanding their meanings—much like "random parrots." This paper sparked a conflict between her and Google's upper management, ultimately leading to her forced departure.

Gebru's experience marks the entry of AI development into a new stage. When AI is no longer just a toy in the laboratory but a powerful tool capable of influencing hiring, credit approvals, and even judicial decisions, examining its biases, risks, and social impacts becomes crucial. Her work serves as a reminder to the entire industry: if technological progress is divorced from humanistic care and social responsibility, it may bring not welfare but new shackles. From Goodfellow's creation to Ng's popularization and Gebru's reflection, the stories of these three figures collectively outline the complete picture of AI in the new era: a technology with infinite creativity is integrating into the world at an unprecedented speed, while also forcing us to seriously consider how to coexist with it.

Starting with a Question

"Can machines think?" Every great transformation often does not begin with a groundbreaking invention but stems from a shocking question. The genesis of artificial intelligence (AI) is just that. Its starting point is not a specific machine or a line of magical code, but a philosophical inquiry posed to the world by British mathematician Alan Turing in the mid-20th century: "Can machines think?" In an era when computers were as large as rooms, this question sounded like science fiction. But Turing's extraordinary insight was that he did not let this question remain in philosophical speculation. He designed a clever thought experiment—the "Imitation Game," which later became widely known as the "Turing Test." This test cleverly sidestepped the definition of the vague concept of "thinking" and instead proposed: if a machine can converse with humans and its performance makes it indistinguishable from a human, can we then consider this machine to possess intelligence?

The proposal of this question was like a lightning bolt that pierced the long night. It brought the ancient dream of "creating intelligence" from the realms of myth and philosophy into a verifiable and challengeable engineering domain. Turing did not provide us with answers, but he gave all successors a clear goal and a blueprint to start drawing. He told the world: intelligence may be precisely describable and simulatable. This spark of thought was sown in a piece of extremely fertile soil.

The post-World War II world, especially in the 1950s, was filled with an unprecedented spirit of scientific optimism and a "can-do" attitude. Humanity had just harnessed atomic energy, invented electronic computers, and deciphered the code of life. The victories of science led people to believe that with human wisdom and powerful new tools, no grand challenge was insurmountable. If machines could calculate complex ballistics and crack enemy codes, why couldn't they go further to mimic or even possess human learning, reasoning, and creative abilities?

It was against this backdrop that a group of the most brilliant and visionary minds of the time began to be drawn to the same dream. However, their sparks of thought were scattered across various fields, including mathematics, psychology, information theory, and the emerging field of computer science. They needed an opportunity, a moment to converge these scattered streams into a mighty river. That moment arrived in the summer of 1956. A young mathematician named John McCarthy, along with Marvin Minsky, Nathaniel Rochester, and Claude Shannon, the father of information theory, submitted a bold proposal to the Rockefeller Foundation.

They planned to hold a several-week-long summer seminar at Dartmouth College in New Hampshire, USA. The opening of the proposal was filled with the optimism and ambition of that era: "We propose to conduct research on artificial intelligence in the summer of 1956... based on the conjecture that every aspect of learning or any other feature of intelligence can, in principle, be precisely described, allowing machines to simulate it." To give this new field a clear identity, McCarthy painstakingly created a new term: "Artificial Intelligence."

This choice was not accidental. At that time, a field called "Cybernetics" was already influential, primarily studying feedback and control systems in biology and machines. But McCarthy wanted to carve out a new direction, one focused on achieving advanced cognitive functions like logic and reasoning through computers, rather than being constrained by the framework of Cybernetics. The birth of this name was like a resounding "declaration of independence," providing a common banner for all explorers with the same dream.

In the summer of 1956, this gathering, later known as the "Dartmouth Conference," took place as scheduled. It was not a rigorous academic conference but rather a long brainstorming session lasting six to eight weeks. Allen Newell, Herbert Simon, Ray Solomonoff, and other future giants of the AI field were all present. They came from diverse backgrounds, bringing different perspectives from logic, psychology, mathematics, and engineering.

One of the highlights of the conference was the demonstration of the "Logic Theorist" program by Newell and Simon. This program successfully proved several theorems from the famous mathematical work "Principia Mathematica." This was not just a technical demonstration; it was a declaration to the world that machines could indeed perform complex symbolic reasoning tasks once thought to be exclusive to humans. It provided the first affirmative, visible answer to the question, "Can machines think?"

This is the "Genesis" of AI. It did not emerge from a one-time success in a laboratory but occurred within a great convergence of ideas. The importance of the Dartmouth Conference can be summarized in three points: First, it named the field. From then on, "Artificial Intelligence" had an official identity, attracting subsequent funding, talent, and attention. Second, it established the core agenda. The conference explored directions such as symbolic processing, neural networks, and natural language processing, which became the main avenues of AI research in the following decades. Third, it built the initial community. This conference connected a group of solitary thinkers into an academic community, and upon returning to their respective institutions, they established the earliest AI labs (such as MIT, Carnegie Mellon University, and Stanford University), sowing the seeds for future towering trees. The Dartmouth Conference is hailed by later generations as the "Constitutional Convention of AI." It officially transformed Turing's great question into a grand scientific journey that attracted generations of top talent.

Although the attendees were overly optimistic in their predictions for the future and failed to foresee the challenges and "winters" ahead, the flame they ignited never extinguished. From a question to the birth of a discipline, the story of AI has begun.

The Genius's Fantasies and the Wall of Reality

After the Dartmouth Conference officially named artificial intelligence, a "golden age" (approximately 1956-1974) filled with boundless optimism and bold predictions began.

These early AI pioneers, represented by Herbert Simon and Marvin Minsky, firmly believed they had mastered the key to machine intelligence. Their confidence was not unfounded; it was ignited by a series of astonishing successes achieved in "miniature worlds." The most representative of these early achievements was the "Logic Theorist" program. Developed in 1956 by Allen Newell, Herbert Simon, and J.C. Shaw, this program is widely regarded as the world's first artificial intelligence program.

Its task was to prove the mathematical theorems proposed by mathematicians Whitehead and Russell in their monumental work "Principia Mathematica." The results were shocking: the "Logic Theorist" not only successfully proved 38 of the first 52 theorems in the book but even found a more concise and elegant proof for one of them than the original text. This achievement greatly encouraged researchers, as it clearly demonstrated that machines could not only compute but also engage in complex logical reasoning activities once thought to be exclusive to humans.

Following this, the team launched the "General Problem Solver" (GPS) in 1959. The revolutionary aspect of GPS was that it attempted to simulate the general thinking process humans use to solve problems. It separated specific domain knowledge (like rules) from general solving strategies, employing a strategy called "means-ends analysis" to continuously set sub-goals to approach the final answer. GPS successfully solved a series of classic logical puzzles, such as the Tower of Hanoi and geometric proofs, giving people hope for creating a "thinking machine" capable of solving general problems across domains.

If GPS demonstrated the machine's "thinking" ability, then MIT's SHRDLU system allowed machines to possess the "understanding" ability to interact with the physical world for the first time. In this virtual "block world" created by Terry Winograd in 1970, users could issue commands to the system in everyday English, such as "Pick up that big red block." SHRDLU could parse commands, understand context (for example, when you ask "Which pyramid?" it would proactively request clarification), plan and execute a series of actions (like grabbing, moving, stacking), and even answer questions about the state of this world. The success of SHRDLU perfectly integrated language understanding, reasoning, planning, and action execution, making it seem as if intelligent robots capable of free dialogue and collaboration with humans were emerging from science fiction films. These brilliant victories achieved in closed, rule-defined "toy worlds" fostered immense optimism.

Simon boldly predicted in 1965, "Within twenty years, machines will be able to do everything that humans can do." Minsky echoed, "In a generation... the problem of creating 'artificial intelligence' will be essentially solved." However, when these geniuses' fantasies attempted to transition from idealized laboratories to the complex realities of the world, they quickly hit a hard and cold wall. This wall was composed of two fundamental problems. The first was "combinatorial explosion."

In simple block worlds, possibilities are limited. But when the scale of the problem expands slightly, such as moving from checkers to Go, or from planning the movement of a few blocks to planning city traffic, the number of possibilities that need to be computed can grow exponentially, instantly exhausting even the most powerful computers of the time and today. The elegance of early AI in "toy problems" became utterly vulnerable in the face of the complexities of reality. The second problem was even more fundamental—the "lack of common sense and context."

The human world is filled with a vast amount of self-evident common sense and ambiguous context. For example, we know that "water is wet," "a rope can be pulled but not pushed," and "if a person gets caught in the rain, they might catch a cold." This knowledge is so basic that we often do not even realize its existence. But for an AI system that only understands logic and rules, this world is completely foreign. It cannot comprehend these implicit background knowledge, leading to its reasoning abilities appearing extremely weak and absurd in real-world scenarios.

SHRDLU could understand "pick up a block," but it could not understand what "pick up a promise" meant. This dilemma of "symbol grounding," where symbols cannot be associated with the meanings of the real world, became an insurmountable chasm for symbolic AI. The gap between high expectations and harsh realities created a significant disparity, and disappointment began to spread, ultimately igniting the first "winter" of artificial intelligence through two landmark events. The first event was the publication of the "Lighthill Report" by the British government in 1973.

This report, written by applied mathematician Sir James Lighthill, delivered a ruthless critique of the AI research of the time. The report sharply pointed out that AI had "yet to achieve any significant impact in automation and language processing." It directly addressed the two major weaknesses of AI research: encountering "combinatorial explosion" when solving real-world problems and being completely unable to handle "common sense." This highly influential report directly led to significant cuts in funding for university AI research by the British government, causing AI research in the UK to nearly come to a halt.

The second heavy blow came from the United States, delivered by none other than Marvin Minsky himself, a leading figure in the AI field. In 1969, Minsky co-authored the book "Perceptrons" with Seymour Papert. In this book, they rigorously proved the fundamental limitations of another technical route parallel to symbolism—connectionism (the precursor to neural networks). They demonstrated that single-layer neural networks (i.e., "perceptrons") are linear models that cannot solve some basic problems, such as the simplest "XOR" problem.

This conclusion was correct in itself, but it was interpreted by the outside world as a "death sentence" for the entire neural network route. The immense influence of this book led to a near-total halt in funding for connectionism research, causing this path, which could have complemented symbolism, to enter a prolonged silence of over a decade. Thus, the once fervent enthusiasm quickly cooled. The high expectations, insurmountable theoretical bottlenecks, and the subsequent withdrawal of funding collectively pushed artificial intelligence into its first long winter.

The fantasies of geniuses collided with the wall of reality, and the entire field fell from the noisy peak into the silent valley, waiting for the next revival in stealth.

Stealth and Revival

Finding a Way Out in Uncertainty In the late 1980s, the "golden age" of artificial intelligence faced a biting cold wind. The once-promising expert systems market collapsed, the LISP machine industry declined, and government and corporate investment enthusiasm sharply cooled. AI research once again entered a low point, marking the second "AI winter" in history.

However, unlike the near silence of the first winter, this time, AI development did not come to a complete halt; it was like a frozen river, seemingly quiet on the surface, but two undercurrents were quietly surging beneath. One was the "mainstream" striving to prove its value in specific domains, while the other was the "undercurrent" silently accumulating strength, waiting for spring. The first route was the survival exploration of symbolic AI in adversity. Although expert systems ultimately declined due to high costs of knowledge base construction and difficulties in handling uncertainty, they left behind a valuable legacy: they proved that AI could solve real problems in specific scenarios, lighting the first lamp for the commercialization of AI.

More importantly, in reflecting on why expert systems failed, a thinker pointed out a whole new direction for AI development. He is Judea Pearl. Pearl realized that the real world is filled with uncertainty, and black-and-white logical rules are far from sufficient to describe the complexity of the world. He introduced probability theory and causal inference into AI, teaching machines how to think in terms of "possibilities" and how to make reasonable decisions in the face of incomplete information.

This was not only an important supplement to symbolism but also a crucial step for AI to move from an idealized logical world to a real world filled with unknowns and changes. Meanwhile, another more hidden and revolutionary route was "stealthily" developing on the fringes of academia. This was the study of connectionism represented by neural networks. The explorers of this route were the true "deep divers." They added a powerful theoretical weapon to their arsenal. In 1986, Geoffrey Hinton and his colleagues reintroduced the backpropagation algorithm and systematically proved its effectiveness.

This algorithm cleverly solved the training problem of multi-layer neural networks, allowing machines to adjust their internal parameters layer by layer through "reflecting" on errors, thus learning more complex patterns. Hinton later recalled that they optimistically believed this algorithm "would solve everything." However, the theoretical dawn did not immediately dispel the cold winter of reality. Entering the 1990s, neural network research quickly hit three high walls: insufficient computing power, data scarcity, and academic skepticism from peers. The computing performance of the time was weak and could not support the training of large-scale networks.

At the same time, statistical learning methods like Support Vector Machines (SVMs) performed better and more efficiently than neural networks on many tasks, causing a significant flow of research funding and talent to other fields. Neural networks were once again viewed as impractical dragon-slaying techniques, and research fell into a predicament of funding difficulties and neglect. In such a challenging environment, some researchers chose to persevere. Yann LeCun was one of the most outstanding representatives.

In 1988, he joined AT&T Bell Labs and, under pressure from the mainstream academic community, devoted all his energy to developing a special type of neural network—convolutional neural networks (CNNs). He firmly believed that this network structure, which mimicked the biological visual cortex, was the key to enabling machines to "see" the world. LeCun's goal was very clear: to enable machines to recognize handwritten bank checks. After years of iteration, he led his team to launch the classic LeNet-5 model in 1998.

This network was successfully deployed in commercial systems, processing approximately 20 million checks daily by the early 21st century, accounting for about 10% of the total check circulation in the United States at the time. This was a milestone success. It was not only a rare commercial victory for neural network technology during the winter but also like a seed buried in frozen soil, proving the immense energy contained in this "undercurrent." It told the world: neural networks are not a fantasy; they can solve real-world problems and have unlimited potential.

Thus, from the 1990s to the early 21st century, the two routes of AI developed in parallel. One route, using probability and causality as tools, allowed AI to "survive" in the business world and learn to coexist more maturely with uncertainty; the other route "stealthily" honed the sharpest weapons for the upcoming revolution under the perseverance of a few. These two forces, one in the light and one in the dark, together laid the groundwork for the impending explosion.

At this time, a "tailwind" from the hardware field had also quietly begun to rise—the parallel computing hardware represented by GPUs, with its powerful matrix computation capabilities, naturally aligned with the computational demands of neural networks. When this hardware tailwind finally blew into the fertile ground of computing power, data, and algorithms, a technological revolution that would change the world was about to unfold.

From "Understanding the World" to "Creating the World"

On September 30, 2012, a historic turning point arrived at the ImageNet Large Scale Visual Recognition Challenge (ILSVRC).

A team composed of Professor Geoffrey Hinton and his two students—Alex Krizhevsky and Ilya Sutskever—submitted a deep neural network model called AlexNet. Its performance shocked the entire computer vision field: its image recognition error rate was only 15.3%, a full 10.8 percentage points lower than the second-place team.

This was not just a victory in a competition; it was a starting gun. AlexNet irrefutably demonstrated that with deep networks, massive data, and the powerful computing capabilities of GPUs, machines could indeed learn to "understand" this world. From then on, the revolution of deep learning was thoroughly ignited, and the development of AI entered a new era. If AlexNet gave AI an unprecedented "eye," then just two years later, a young researcher endowed AI with a boundless "imagination."

In 2014, while still a PhD student, Ian Goodfellow had a flash of inspiration during a discussion with friends at a bar, proposing a genius concept—Generative Adversarial Networks (GANs). The principle of GAN is like an eternal competition between "spear" and "shield." It consists of two neural networks that play against each other: a "generator" and a "discriminator."

The generator's task is to create data (like images) that are indistinguishable from real data, striving to deceive the discriminator; while the discriminator's task is to discern which data is real and which is fabricated by the generator. In this ongoing confrontation and evolution, the generator's "forgery" skills become increasingly sophisticated, ultimately enabling it to create new content that is difficult for humans to distinguish. From high-definition faces to artistic paintings and medical images, GAN allowed AI to transform from a "recognizer" and "analyzer" into a "creator" for the first time.

AI was no longer just about understanding the world; it began to have the ability to create a brand new, digital "world." Just as AI was making great strides in visual creation, a more profound structural change was quietly brewing. In 2017, a research team from Google published a groundbreaking paper titled "Attention Is All You Need." This paper abandoned the recurrent neural network (RNN) structure that had been commonly relied upon for processing sequential data (like language) and proposed a completely new architecture—Transformer.

The core of the Transformer is a design called "self-attention mechanism," which not only better captures long-distance dependencies in text but, more crucially, enables efficient parallel computation, greatly enhancing the training speed and scalability of the model. The birth of the Transformer architecture laid a solid foundation for AI, paving the way for a series of large language models (LLMs) to explode.

Starting with the first GPT model released by OpenAI in 2018, this technological route rapidly iterated. The model's parameter count and data scale grew exponentially, and AI's capabilities underwent qualitative changes, evolving from simple text generation to being able to engage in fluent conversations, write code, and even exhibit astonishing abilities like "few-shot learning" with GPT-3. AI's creativity extended from images to the most core domain of human intelligence—language.

The Diffusion of Power and Reflection

From Engineering to Ethical Constraints Technological breakthroughs, to change the world, require overcoming the "last mile" from the laboratory to industry.

In this process, figures like Andrew Ng played a key "evangelist" role. They dedicated themselves to promoting the engineering and popularization of AI, transforming complex deep learning technologies into scalable tools and courses, enabling thousands of engineers and learners to master and apply AI, significantly accelerating the diffusion of AI power across various sectors of society. However, when a power becomes strong enough, it brings not only opportunities but also risks.

In 2019, when OpenAI released its new model GPT-2, it took an unprecedented cautious approach. Concerned that its powerful text generation capabilities could be used for malicious purposes such as creating fake news, spam, or online bullying, they initially chose to release only a small version and withheld the full model. This move sparked intense debates in the tech community about "open research" versus "responsible disclosure." Ultimately, after observing "no strong evidence of misuse," OpenAI publicly released the complete 1.5B parameter model in November of the same year. The controversy surrounding the release of GPT-2 was just the tip of the iceberg. With the proliferation of generative AI capabilities, Deepfake technology began to emerge as a serious social issue. Using AI for face-swapping or voice synthesis could easily create fake videos or audio of political figures, spreading misinformation, undermining public trust, and even interfering with election processes. Moreover, researchers soon discovered that these models, trained on vast amounts of internet data, faithfully reflected the biases present in human society.

For example, analyses showed that GPT-2, when describing professions, would unconsciously associate women with more stereotypical jobs. In the face of these challenges, a new, critical voice began to resonate in the AI field. AI ethics researchers, represented by Timnit Gebru, began to call out: in the pursuit of stronger models, we must seriously examine the social impacts, algorithmic biases, and potential risks of technology. The research directions they promoted, from how to detect and mitigate model biases to how to establish responsible AI governance frameworks, marked the entry of AI development into a new stage that requires serious societal consideration and constraints.

From AlexNet enabling AI to "understand" the world in 2012, to GANs and Transformers allowing AI to "create" the world, and now to the necessity of contemplating how to "constrain" this increasingly powerful AI. This leap in just over a decade is not only a leap in technological capability but also a profound reshaping of the relationship between AI and human society. AI is no longer just a tool in the engineer's study; it has become a powerful force shaping our reality and influencing our future.

In Conclusion: What Will the Next "Turing" Ask?

Looking back at the journey of artificial intelligence over the past seventy years, we see not a "invention" sparked by a genius but a relay race of ideas spanning generations. This long race began with Alan Turing firing the starting gun at the starting line; he did not create AI but defined the finish line of the entire track with a simple yet profound question—"Can machines think?" The baton was first passed to John McCarthy, who officially "named" this emerging field as artificial intelligence at the Dartmouth Conference in 1956, providing the pioneers with a shared identity and banner.

Subsequently, the symbolists, represented by Marvin Minsky, confidently surged ahead, believing that intelligence could be constructed with logic and rules, achieving brilliant early successes. However, the complexities of the real world soon led them to hit a wall, and AI subsequently faced its first winter. In the long silence, the "Three Giants"—Geoffrey Hinton, Yann LeCun, and Yoshua Bengio—quietly guarded the flame of connectionism in the corners where no one was paying attention, firmly believing that neural networks mimicking the brain were the correct path.

Their perseverance ultimately awaited the tailwind of computing power and data. When Ian Goodfellow's Generative Adversarial Networks (GANs) emerged, AI transformed from merely a recognizer into a creator, ushering in the grand era of generative AI. Just as technology was advancing rapidly, the warnings from figures like Timnit Gebru represented the entry of the race into a new stage—we must begin to examine the ethical and social responsibilities of this powerful force. From the historical context, we can see the contours of the future more clearly.

For instance, the early rivalry between symbolism and connectionism has not ended with the victory of deep learning. Today, they are merging in the form of "neuro-symbolic AI," allowing large models to possess stronger logical reasoning abilities and interpretability alongside powerful perceptual capabilities. Similarly, the prominence of AI ethics is not coincidental; it is an inevitable requirement of social development when technological power reaches a critical point. When AI begins to influence employment, shape public opinion, and even participate in historical narratives, discussing its fairness, transparency, and social responsibility becomes a task we must complete.

So, what does understanding this history, composed of countless collisions of ideas, route struggles, and individual perseverance, mean for us ordinary people? The answer is: it helps us build a cognitive framework, dispelling the mystique and anxiety surrounding AI. When we realize that AI did not emerge as a "black technology" out of thin air but originated from a question posed by Turing, a setback experienced by Minsky, and decades of persistence by Hinton, we can view its capabilities and limitations more calmly and think more rationally about its place in our work and lives, rather than being passively swept along by the tide.

The baton of history is now in the hands of our generation. Turing's question defined whether AI "can or cannot," while today, the questions we face may be "should or shouldn't" and "how to coexist." So, what new questions will the next "Turing" ask? Is it about the nature of consciousness, the rights of machines, or humanity's new role after the explosion of intelligence? This question has no standard answer. It hangs in the future, waiting for each of us who are witnessing, participating in, and being influenced by this transformation to think and respond together.