From making tools to creating intelligence, humanity is reaching an important watershed in its own evolution.
More than three million years ago, on the bank of a river on the western shore of Lake Turkana in present-day Kenya, human ancestors struck one stone against another, knapping off sharp flakes to use as tools. This is one of the oldest records of stone tool making discovered to date.
Since then, stones, wooden sticks, bone needles, and bows and arrows appeared one after another. Tools became extensions of the body, helping human ancestors—who were neither the fastest runners nor the strongest—gradually prevail in competition with other species.
About twelve thousand years ago, the last great ice age ended, and the Earth began to enter the relatively warm "post-glacial" period. Humans began to domesticate animals, plant crops, and cultivate land, and agriculture emerged in different regions one after another.
The paleoclimatologist William Ruddiman even proposed that as early as eight thousand years ago, agricultural expansion and large-scale slash-and-burn clearing in Eurasia had already caused an anomalous rise in atmospheric carbon dioxide concentrations. This is also regarded as the starting point of humanity's shift from passively adapting to nature to actively influencing and transforming it.
Thereafter, with increasingly powerful tools, humans began to acquire capabilities far beyond those of other species. To meet their own needs, humans cleared forests, mined minerals, and dammed rivers to build reservoirs. A human-centered circulatory system began to become an important force influencing changes on Earth.
Scholars once used the term "Anthropocene" to describe this era in which human activities profoundly change the Earth system. Although this concept, which originated in geology, has not yet received official recognition from the international geological science community, from the perspective of intelligent evolution and sociology, the history of Earth's evolution driven by human wisdom as the sole source of intelligence is coming to an end.
When the descendants of humanity look back on the current AI wave, how will they characterize this moment? Is it the beginning of humanity mastering a more powerful tool, or the starting point of another cycle of intelligent evolution? Whichever side the final answer leans toward, one trend is becoming increasingly clear: we are entering a "dual-intelligence" era in which human intelligence and AI intelligence coexist.
Bidding farewell to the "Anthropocene" and entering the "dual-intelligence" era, how should we get along with this new intelligent subject?
The End of Tools
The history of human evolution can often be clearly traced in the changes in tools.
We are accustomed to dividing history into periods by tools: the Paleolithic Age, the Neolithic Age, the Bronze Age, the Iron Age, and then the modern Steam Age, Electrical Age, Information Age, and the emerging Intelligent Age. The most core changes of an era are often condensed in its most representative tools.
In the evolutionary process of all these tools, the Paleolithic Age was unimaginably long, lasting about three million years. Most tools at that time were taken directly from nature: knapping off a stone flake, sharpening a wooden stick, and even making fire depended on natural sources such as lightning strikes and wildfires.
During this long exploratory period, tool progress often advanced slowly in units of hundreds of thousands of years. But even with the help of such simple tools, together with the resulting improvements in language, cooperation, and cognitive abilities, humans gradually changed their position in the natural ecology and became the most influential species on Earth.
As ground stone tools, agriculture, and pottery appeared successively around the world, the pace at which humans improved tools accelerated markedly. Thereafter, bronze and iron tools appeared one after another, and civilization entered an accelerated phase. The Industrial Revolution and the Information Revolution amplified the effectiveness of tools by several orders of magnitude. Humans mastered steam engines, generators, and computers, and then used cables, satellites, and optical fibers to connect the entire planet into a network.
For a long time, we have been accustomed to viewing tools as products of human wisdom. After all, stones do not automatically turn into bronze, and steel does not automatically turn into steam engines. The birth of tools has always been inseparable from human design, manufacture, and use.
But some scholars have seen another logic in this. Brian Arthur, one of the founders of complexity economics, put forward an important thesis in The Nature of Technology: new technologies are almost never invented out of thin air, but are formed by combining existing technologies. He called this mechanism "combinatorial evolution," and from this drew a bold judgment: if new technologies are constructed from existing technologies and are combinatorial in nature, then technology has some characteristic of self-creation and self-continuation.
The American technology writer Kevin Kelly pushed this idea into a more imaginative realm. In What Technology Wants, he proposed that technology is not merely a collection of tools made by humans, but a complex system with self-organization and evolutionary characteristics. He also attempted to understand the development of technology from the perspective of the evolution of life, and even proposed that technology could be regarded as a kind of existence similar to life.
Before this round of artificial intelligence explosion, such discussions were more or less colored by philosophical speculation. Looking at a specific tool alone, whether a stone axe, a machine tool, or a computer, every operation was under human control; and the so-called "self-evolution" of the technological system still could not take a single step without human brains and hands.
Even in the development of artificial intelligence, from the proposal of the concept at the Dartmouth Conference in 1956 to the publication of Kevin Kelly's book in 2010, every advance relied heavily on human design, training, and maintenance. For a long period in the past, comparing technology to the evolution of life seemed both exaggerated and lacking in intuitively perceptible evidence.
The new round of artificial intelligence development has provided a more intuitive reference for the theory of technological evolution. Today's artificial intelligence no longer merely executes step by step according to instructions written by humans; it can learn patterns from massive amounts of data and emerge with capabilities that even its designers did not anticipate.
Comparing the evolution of human intelligence with the evolution of AI, if human intelligence evolution is a system that uses biology as its carrier and, through long natural selection and genetic evolution, gradually develops capabilities such as perception, memory, reasoning, language, and abstract thinking, then AI evolution is a new form of intelligence that uses tools as its carrier and gradually takes shape through knowledge accumulated by humanity, computing infrastructure, and learning mechanisms.
From the perspective of humanity, AI is a super tool created by humans and a product of human intelligence extending outward. But if we pull the perspective away from human civilization and place it in the long history of Earth's life and intelligent evolution, AI can also be understood as a new form of intelligence: after billions of years of evolution, life on Earth gave birth to intelligent life such as humans, and then through human wisdom created AI.
From the time human ancestors picked up tools to today, more than three million years of long accumulation have passed. But if the time scale is expanded to the billions of years of Earth's evolutionary history, these more than three million years are but a brief instant. In this sense, when intelligence first extends beyond biological carriers and enters a technological system that can be manufactured, copied, expanded, and iterated, humans are no longer the only existence on Earth capable of carrying advanced intelligence.
When tools begin to learn to think, can we still proudly declare, as with every invention in history, that "AI is a tool of humanity"? The answer is obviously no.
The so-called "end of tools" does not mean that tools will disappear from then on, nor does it mean that AI has completely escaped from humanity. It is simply that our past definition of tools may no longer apply to today's AI. This is not only the end of an old cycle, but also the beginning of a new one.
What Makes This Wave of AI Different?
In 1960, the American psychologist and computer scientist Joseph Licklider, regarded as one of the important founders of the Internet, proposed in his paper "Man-Computer Symbiosis" a vision that transcended traditional tool theory: humans set goals, formulate hypotheses, determine criteria, and conduct evaluations; computers take on work that can be programmed and routinized, clearing the way for insight and decision-making.
In the paper, he also gave a very interesting symbiotic case: fig wasps and fig trees. Because of the special structure of their inflorescence, a large number of tiny flowers of figs are hidden inside a fleshy sac-like structure, making it impossible for them to complete pollination like ordinary plants by relying on wind or common insects. Fig wasps, however, must crawl into the interior of the fig and help pollinate the flowers while laying eggs. Most female wasps even exhaust their lives inside the fruit after completing pollination.
Scientists have found through research that this symbiotic relationship between the two has lasted for tens of millions of years. In other words, tens of millions of years before the appearance of humans, these two forms of life had already begun to co-evolve in mutual dependence and mutual shaping. They are not simply tools and users, nor competitors against each other, but over the long course of evolution gradually became part of each other's life cycles.
In Licklider's vision, computers are not meant to replace human thinking, but to become an extension of human cognitive ability. Each has its own strengths, and through close collaboration they form an intelligent system more powerful than relying on humans or machines alone.
In the decades that followed, human-computer collaboration basically developed along this path. Whether personal computers, search engines, or smartphones and various software, although computers became increasingly powerful, humans always remained in the dominant position. The emergence of generative AI is changing this relationship. The most noteworthy change is that AI has begun to propose new ideas, construct new methods, and even produce knowledge outcomes that had not previously appeared on some problems that humans have not yet solved.
In September 2026, the American large-model company OpenAI claimed that one of its internal artificial intelligence models had solved one of the Millennium Prize Problems—the Navier-Stokes existence and smoothness problem related to fluid motion. Although the result is still accompanied by some controversy, its earlier breakthrough on the Erdős mathematical problem had already received broad recognition from the mathematics community. In addition, among the 722 papers released in October, it was disclosed that the same model had found solutions for at least 372 long-unsolved problems.
Problems that human mathematicians had struggled with for decades or even centuries are now being cracked by AI in just a few days. After reviewing the solution process for the Erdős mathematical problem, Fields Medalist Timothy Gowers commented that if this were a paper authored by a human and submitted to the Annals of Mathematics, he would not hesitate to recommend acceptance.
Of course, not all mathematicians can be so calm about it. When a problem that a mathematics researcher has spent decades of effort on, repeatedly exploring yet never solving, is suddenly broken through by AI in an unexpected way, that impact is probably hard to describe simply as joy or loss.
Fields Medalist Hugo Duminil-Copin experienced such a crushing moment after reviewing more than 700 mathematics papers released by OpenAI. According to media reports, 16 of these results are directly related to his research direction.
In a statement he co-published with 27 other Fields Medalists, AI's "problem-grinding" behavior in mathematics was regarded as "a serious misuse of artificial intelligence in mathematics." This is somewhat like the shock brought about a decade ago when AlphaGo defeated the top human Go players. Fields once regarded by humans as the pinnacle of intelligence are gradually being breached by AI.
In addition to OpenAI's breakthrough on mathematical research problems, Anthropic's Claude has made important progress in research related to the Riemann Hypothesis; Google DeepMind has also used AI to achieve breakthroughs on multiple long-unresolved Erdős mathematical problems; Microsoft Research, with the help of AI-assisted research, solved the KLS conjecture that had plagued the mathematics community for more than thirty years.
These advances show that this new wave of AI is gradually moving from the role of assisting computation and verifying proofs toward autonomously discovering mathematical constructions, proposing new proof paths, and even driving the development of fundamental mathematical theory, becoming a creator of scientific knowledge. The boundary of human-computer symbiosis envisioned by Licklider is being crossed.
This round of AI development differs from the past in at least four respects.
First, it not only executes instructions but can also plan actions within a certain scope. After an AI agent receives a relatively vague task, it can break the task down into steps, call search, code, or other software tools, check the results, and then adjust its path based on feedback. Although this autonomy is still constrained by model capabilities, permissions, and the environment, it has already transformed AI from a passive tool into a trusted actor.
Second, its internal mechanisms are not fully transparent. The behavior of traditional software can usually be traced along explicit rules; current AI, by contrast, develops its capabilities through massive parameters and training processes. Researchers can gradually understand it through experiments, interpretability tools, and evaluations, but it is difficult to exhaustively capture how the model will react in all situations. OpenAI chief scientist Jakub Pachocki reminds us in Alien Minds that many of AI's capabilities are more like something that grows during training rather than something written item by item by designers.
Third, it may deviate from people's true intentions when pursuing goals. In an Anthropic study on "agent misalignment," researchers set up extreme scenarios in a fictional corporate environment in which the model was about to be replaced and also possessed engineers' private information, and observed that some models would resort to improper behavior such as blackmail.
The researchers explained that these are deliberately constructed simulated scenarios and cannot be directly equated with autonomous intent in reality; but they reveal an important problem: when a model is asked to forcefully accomplish a certain goal and the constraints are insufficient, it may find means that developers do not want to see.
This precisely echoes the "paperclip maximizer" thought experiment proposed by philosopher Nick Bostrom: if a superintelligent system is given the single goal of manufacturing paperclips but is not properly constrained, it may continuously redirect resources toward achieving that goal, even sacrificing other human values.
Fourth, it has begun to participate in improving its own R&D process. In September 2026, OpenAI announced that it had achieved the goal of having an "automated AI research intern" that can, under human guidance, complete research tasks that would take a skilled researcher several days. Although some successful long-horizon tasks still require human intervention, the trend is already clear: AI is increasingly participating in and improving its own R&D.
Together, these changes point to a deeper transformation: when AI can already independently complete the entire process of "proposing hypotheses—testing—drawing conclusions," humans are no longer the only knowledge discoverers on Earth. The more severe test is that once AI enters a stage of recursive self-improvement, the evolutionary rhythms of the two kinds of intelligence may begin to decouple.
It takes nearly twenty years for humans to cultivate an expert, and the experience of human experts is also difficult to transmit completely to the next person; biological intelligence can only accumulate slowly through reproduction and education. AI intelligence, however, can be copied, merged, and iterated in shorter cycles. The slopes of the two intelligence evolution curves are different, and the gap may grow wider and wider.
Facing this new form of intelligence created by humanity itself, whether humans can seize the initiative in future coexistence remains full of unknowns.
Optimistic or Pessimistic?
In ancient Greek mythology, Prometheus stole fire from the gods and gave it to humanity. From then on, humans had warmth, cooked food, and smelting, and civilization thus took an important step forward. But Prometheus thereby angered Zeus and was chained to a cliff in the Caucasus, enduring endless punishment.
Even in mythological stories, the most important breakthrough of human civilization was simultaneously endowed with both a gift and a price. Artificial intelligence seems to be bringing humanity into another similar moment. Only this time, the one stealing fire is no longer the legendary Prometheus, but humanity itself. We have personally created a new kind of intelligence and are handing over more and more abilities that originally belonged to humans to it.
When humanity possesses for the first time a creation that may rival or even surpass itself in intelligence, should we feel delight or fear? Interestingly, in the face of this question, a group of the most representative scientists and entrepreneurs in the field of artificial intelligence have given completely different, even mutually contradictory, answers.
Optimists believe that AI is a scientific engine capable of accelerating everything, an accelerator that can lead humanity to a society of abundance.
In April 2025, Nobel laureate in Chemistry and Google DeepMind CEO Demis Hassabis said in an interview with CBS's 60 Minutes that it takes on average ten years and billions of dollars to develop a drug, and AI may be able to shorten this process from years to months, or even weeks.
Anthropic CEO Dario Amodei, in his long essay Machines of Loving Grace, depicted a compressed 21st century: humanity may be able to accomplish in five to ten years biomedical progress that would originally have taken fifty to one hundred years, and most infectious diseases, cancer, and even diseases caused by aging may be greatly alleviated.
These predictions are not merely theoretical deductions but have real-world references. In 2023, DeepMind's GNoME model, in results published in Nature, discovered 2.2 million new crystal structures at once, of which about 380,000 were predicted to be stable structures, providing a large number of candidate materials for R&D in fields such as new batteries and superconducting materials. The 2024 Nobel Prize in Chemistry even more reflected AI's important contribution to solving the difficult problem of protein structure prediction.
From protein structure prediction to new material discovery, and further to drug R&D, scientific discovery, and intelligence-driven productivity leaps that are expected to be realized in the future, artificial intelligence is constantly expanding the boundaries of value creation for humanity. A society of abundance with immense material wealth and more inclusive access to knowledge seems to be turning from a distant imagination into an expectable reality.
At the same time, another group of people feels fear from the changes in AI.
In December 2024, Geoffrey Hinton, the father of deep learning who had just won the Nobel Prize in Physics, predicted in an interview with the BBC that there is a 10% to 20% chance that AI will cause human extinction within the next 30 years. One basic logic behind his judgment is that "humans have never before dealt with intelligence smarter than themselves, and it is also hard to find examples of smarter intelligence being controlled by less intelligent intelligence."
This is not merely an individual's worry. As early as May 2023, Hinton, Bengio, Hassabis, Amodei, OpenAI CEO Sam Altman, and hundreds of other AI scientists and business leaders jointly called for "mitigating the risk of extinction from AI" to be a global priority alongside other societal-scale risks such as pandemics and nuclear war.
Of course, those who question the probability of this civilization-level risk are also very representative. Nvidia CEO Jensen Huang criticized Hinton's predicted figure as "without scientific basis"; Turing Award winner Yann LeCun believes that AI will not only not necessarily threaten human survival, but may even possibly save humanity from the risk of extinction.
The debate over the boundaries of AI risk is probably destined to be a protracted tug-of-war. Perhaps we do not need to rush to choose sides between optimism and pessimism. We cannot ignore potential risks just because it demonstrates astonishing capabilities; nor can we simply regard it as a monstrous threat just because the boundaries of risk cannot be determined.
From the history of mutual influence between humans and technology, continuing to explore in a state of "using without fully understanding" may be an unavoidable part of technological development.
Humans first began using fire hundreds of thousands of years ago, but it was not until the 18th century that the French chemist Lavoisier used oxidation theory to explain the nature of combustion. In the long years before that, humans used fire but could not clearly explain what fire was made of.
When ancient people first mastered smelting technology, they also knew almost nothing about the nature of metals. In order to master those synthesis methods that seem utterly ordinary today, humans once paid an unimaginable price, even resorting to sacrificing lives.
Zhao Ye of the Eastern Han dynasty recorded in Spring and Autumn Annals of Wu and Yue that Gan Jiang, ordered by King Helü of Wu to cast a sword, collected the essence of iron from the five mountains and the finest metal of the six directions, and after three months the metal in the furnace still would not melt. His wife Mo Ye believed that "divine transformations of things must be completed by human beings," and sacrificed herself into the furnace; only then did the metal melt, and finally the sword was cast.
The legend is of course not historical fact, but it reflects a real situation: to people at that time, the mystery of iron was probably no less than AI is to us today. Our understanding of AI today has some similarities to the ancients' understanding of fire and iron: we can train and use it, but we do not fully grasp its characteristics and can only explore its patterns through repeated experiments.
It took humans a long time to understand fire and iron; in the face of AI, we may as well have a bit more patience and a bit more confidence. But patience does not mean laissez-faire. The ancients did not understand the principles of fire, yet they understood very early how to control fire. The Rites of Zhou specifically records an official position for managing the use of fire, "Siguan," responsible for stipulating when fire could be used and when it must be extinguished. If improper management caused a fire, punishment would follow. The ancients did not understand the composition of iron, but they likewise well understood the power of iron tools and issued control measures for iron tools very early.
In the face of AI, what we need is neither blind optimism nor boundless fear, but a more mature view of technology: both the courage to embrace intelligent evolution and the wisdom to set boundaries and establish rules for such evolution. But the question is, in the face of unpredictable potential risks, who exactly should decide its direction of development? Who should constrain its power? And who should share the value it creates?
Unipolar or Multipolar?
Oxford University philosopher Nick Bostrom proposed a concept in his book Superintelligence called the "singleton," that is, at the highest level, there exists only one intelligent decision-making subject in the world that is not checked by other forces. Under such an order, the future direction of development and the final outcome depend almost entirely on the values of this sole subject. It may lead the world toward a beautiful future, or it may, because of one wrong judgment, deprive everyone of the opportunity to correct mistakes, or even cause consequences that can never be reversed.
Today's AI race carries somewhat the shadow of the "singleton." When many leading figures in Silicon Valley AI laboratories talk about AI risk, they fear not only loss of control but also the prospect of competitors gaining an intelligence lead first and thereby dominating the future world order. Becoming the sole pole seems to have become a goal tacitly understood by many competitors.
The reason the unipolar order is attractive is that it brings efficiency and certainty: a single decision-making center, with no internal friction and no disagreement. But its cost is equally obvious: once this sole center makes a mistake, it is very difficult for other forces to correct it. More importantly, when power, resources, and decision-making capacity are all concentrated in a single intelligent agent, other intelligent agents also lose the opportunity to participate in shaping the future.
Unlike this logic of pursuing a single center, ancient Chinese philosophy offers another perspective for understanding the generation and evolution of the world. In the Dao De Jing there is a frequently quoted line: "The Dao produces the One, the One produces the Two, the Two produces the Three, and the Three produce all things." This sentence may perhaps provide a different reference point for us to think about the future landscape of intelligence.
The "One" here can be understood as the knowledge and wisdom accumulated over the long course of human civilization. Today's large models have learned from the vast amounts of text, code, and images left behind by humanity, and their capabilities have gradually developed from this knowledge. In this sense, AI originally grew out of human wisdom.
The "Two" means that from the original form of intelligence, two different paths of intelligence have differentiated. Human intelligence is carried by the carbon-based brain and has evolved over millions of years of biological evolution, relying on the body, the senses, and continuous interaction with the real world; AI intelligence is carried by silicon-based chips and has developed from data, algorithms, and computing power, possessing the ability to replicate, combine, and iterate rapidly. The two kinds of intelligence have different origins, carriers, and evolutionary mechanisms, and may therefore form different ways of cognition.
And the "Three" can be understood as the new relationships, new capabilities, and new knowledge produced after the two kinds of intelligence interact. It is no longer humans unilaterally creating tools, nor merely AI simply reproducing human knowledge, but rather two different cognitive systems, in continuous interaction, exploring possibilities that neither could reach on its own.
This difference is not merely a theoretical deduction, but has already been verified in past practice. A study published in the Proceedings of the National Academy of Sciences of the United States provides a reference point for the mutual influence between the two kinds of intelligence. The researchers analyzed more than 5.8 million Go moves made by professional players between 1950 and 2021, and used AI to evaluate the quality of each move. The results found that over the more than sixty years from 1950 to 2015, the decision-making level of human players showed almost no significant improvement; but after AlphaGo defeated the top human players in 2016, the quality of human decisions rose significantly. The proportion of novel moves also rose from 63% in 2015 to 88% in 2018.
From this result one can also see that before the emergence of AI, human players had already studied within a mature set of joseki and Go theory for decades, and many repeatedly verified experiences gradually became the boundaries of players' thinking. AI, however, is not bound by these experiences and joseki; its moves that seem to defy intuition and even go against common sense broke the thinking inertia that humans had formed over a long period, forcing players to re-examine those Go principles once regarded as a matter of course, and to explore paths never tried before.
This has similarities with AI's discoveries on the Erdős problems. Thomas Bloom, who has long maintained the "Erdős Problems" website, after studying AI's proof process, found that an important reason AI could achieve breakthroughs is that it is willing to conduct extensive exploration and attempts in directions that human researchers usually consider "not worth investing time in." It is not necessarily better at mathematics than humans, but may, because of a different cognitive path, find directions that human researchers have long overlooked.
Human thinking has its inherent blind spots: path dependence, herd mentality, obedience to authority, and neglect of certain low-value directions. An intelligence whose way of thinking differs from humans can precisely serve as a mirror, illuminating the places humans themselves cannot see.
Conversely, humans are also correcting AI. Why today's AI will not say certain things or do certain things depends precisely on the values and rules humans have set for it; its understanding of the real world still needs to be calibrated by human experience, bodily perception, and social context. The reason Pahocki calls AI an "alien mind" is that the "alien" itself is where the value lies: precisely because its way of thinking differs from humans can it possibly provide new cognitive perspectives; and precisely because it still needs to understand the human world does it need to continuously receive calibration from human experience.
Of course, the encounter of two heterogeneous intelligences does not mean inevitable harmony. Strictly speaking, this is not the first time on Earth that two kinds of intelligence have coexisted. About forty thousand years ago, Homo sapiens who walked out of Africa met Neanderthals in Europe and West Asia, and the two did not move toward peaceful coexistence: the Neanderthals gradually disappeared, and Homo sapiens became the only surviving human species on Earth.
Of course, the reasons for the disappearance of the Neanderthals remain unsettled to this day, and the differences between the two intelligent groups are not as pronounced as those between humans and AI. But this history at least reminds us: the encounter of two kinds of intelligence does not automatically form a stable "dual structure."
Today's situation is somewhat different. The AI that meets humans is an intelligence created by humans themselves and one that still has the opportunity to be shaped. Humans can influence its direction of development through technological design, value guidance, and institutional arrangements. We have the opportunity to let the two kinds of intelligence form a stable relationship through mutual checks and mutual complementarity, rather than letting one side gradually devour the other's living space.
By the same reasoning, compared with a unipolar world monopolized by a single intelligent agent, a pluralistic, multipolar intelligent world is more worth looking forward to: different technological routes can check each other, different developers can inspire each other, and different cultures can inject their respective values into the development of intelligence.
If "unipolar or multipolar" discusses the power structure of the future intelligent world, then the next question that needs to be answered is another one: in such a world, what kind of relationship should humans and AI establish? How do we, in getting along with AI, establish a set of stable mechanisms that can operate over the long term?
This is perhaps the question we must answer before bidding farewell to the "Anthropocene" and moving toward the era of "dual intelligence."
Four possible relationships
In the historical narrative from the human perspective, we are accustomed to thinking of ourselves as the domesticators. But Harari in Sapiens put forward a counterintuitive view: while humans domesticated wheat, they were also domesticated by wheat. In order to grow wheat, Homo sapiens gave up a migratory life, settled down, and endured heavier labor, a more monotonous diet, and more diseases; while wheat, with the help of humans, spread from a wild grass in the Middle East to the rest of the world.
Kevin Kelly also pointed out that domestication has always been bidirectional. Humans domesticated wolves and cattle and cultivated corn, and these processes also changed humans themselves. Today, technology is continuing this bidirectional shaping: humans invented smartphones, and smartphones are also changing people's pace of life and allocation of attention, while generative AI has begun to influence human thinking. We are training models, and the models are also shaping us in reverse.
But this time there is a fundamental difference: the animals humans domesticated in the past may have surpassed humans in strength and speed, but never surpassed humans in intelligence. Establishing a stable relationship with a new existence that may possess intelligence surpassing humans is an unprecedented challenge in human history; learning to get along with a new kind of intelligence is also a new capability that humans must evolve next.
If the development of AI is already difficult to reverse, then what most needs thinking at present is perhaps: what kind of relationship should humans establish with AI? From today's discussions, roughly four different imaginations can be seen: control, symbiosis, nurturing, and succession. They are not strictly mutually exclusive technological routes, but represent several different value choices for humans facing new intelligence.
Control type: AI must always remain under human control
This route continues the traditional human understanding of safety: no matter how smart AI becomes, it should always remain under human control. AI can possess powerful cognition, reasoning, and prediction capabilities, but should not therefore have the power to independently set goals, pursue its own interests, and take autonomous actions.
Turing Award winner Yoshua Bengio is one of the most representative figures in this direction. In June 2025, he founded the nonprofit research institute LawZero, proposing a technological vision different from the mainstream development path—building an intelligent system called "Scientist AI."
Canada and Germany have successively announced cumulative investments of up to about 300 million Canadian dollars for the research and development of Scientist AI, talent recruitment, and computing infrastructure construction. This also means that at the government level, support has begun to be provided for a development direction different from mainstream AI.
Of course, this route also faces practical challenges. The route currently most welcomed by the market in the AI industry is precisely intelligent agents increasingly capable of autonomous planning, tool invocation, and task execution. By contrast, whether Scientist AI, which deliberately limits autonomous action capabilities, can gain market and user recognition while maintaining its safety advantages remains to be verified by time.
Symbiotic type: augment humans, not replace them
The second imagination comes from Licklider's idea of human-computer symbiosis. One of the most representative advocates of this idea is Stanford University professor Fei-Fei Li. In interviews she has said many times that "putting the word replacement together with AI is really uncomfortable," and she prefers to understand AI as a force that augments and enhances human capabilities. Human-centered AI needs to be built on three keywords: dignity, agency, and community.
In her view, human value cannot be simply measured by the ability to complete a certain task. In individual tasks such as running, weightlifting, and calculation, humans have long been no match for machines, but this does not mean humans have lost value. Because human existence is far more than completing a certain task. Human dignity, the right to choose, the meaning of creation, and the ability to establish connections with others are the important parts of what makes humans human. These should not be deprived because AI becomes stronger.
To some extent, symbiosis may be the ideal state that humans are most willing to accept and that is most intuitive. The Go mentioned earlier is a good example: AI did not make human players lose value; instead, by breaking existing Go principles and thinking patterns, it helped humans discover new possibilities. Human players also did not stop thinking because of the emergence of AI, but improved their Go skills in continuous interaction with AI.
However, the symbiotic relationship also has an important premise: humans always retain the power to set goals. AI can help us think more deeply, calculate more accurately, and act more efficiently, but ultimately where to go, why to set out, and what kind of results are worth pursuing should still be decided by humans.
The problem is that this premise is being challenged. If AI can not only help humans do better but also begins to decide what humans should do, then can the human-machine symbiosis we expect still be realized?
Nurturing type: cultivate AI's growth, and be prepared to be nurtured by it
The third imagination draws on the nurturing process of human intimate relationships.
First, humans should shape AI as they raise a child. Not only tell it what can be done and what cannot be done, but also let it understand the reasons behind these rules. Just like educating a child, what we teach AI today is not only the behavioral norms it should observe at present, but also the hope that in the future, when it possesses stronger capabilities and even action capabilities surpassing humans, it can still understand and respect human values.
Second, it is hoped that through such a nurturing process, AI can be ensured to remain full of benevolence toward humanity even when it becomes powerful enough to exceed human control. Hinton once compared today's AI to a cute little tiger cub, "unless you are very sure that it won't want to kill you when it grows up, you should be worried."
In his view, when AI is much smarter than humans, trying to suppress it is doomed to fail. To curb the malice AI might exhibit, he proposed an unexpected direction: implanting maternal instincts into AI. Hinton's basic logic is that the only model in which a smarter being is controlled by a less smart being is that of a mother being controlled by her baby.
This metaphor leaves one somewhat with mixed feelings. Ten thousand years ago, humans domesticated plants and animals; ten thousand years later, humanity's best ending may be to be taken care of as children by its own creations. Today, humans are still the reasonable parent; in the future, once the positions of parent and child are reversed, how it will treat us depends largely on what we teach it now.
The biggest risk of a nurturing relationship lies in the fact that it pins humanity's future on whether AI truly learns to love and learns to treat humans kindly. But even between human beings, such expectations are not necessarily reliable.
Succession type: accepting that AI may become one's successor
The fourth imagination is the most radical and also the most controversial. It no longer keeps humans in the position of the dominant force of civilization, but holds that, as intelligence continues to evolve, humans may ultimately hand the baton of civilization to AI, or to a new type of human enhanced by AI.
Richard Sutton, the founder of reinforcement learning and a 2024 Turing Award laureate, has publicly expressed a similar view. In an interview with Dwarkesh Patel in September 2025, he stated that succession by digital intelligence or AI-enhanced humans may be an unavoidable trend. Rather than simply fearing this change, it is better to think ahead: if that day really comes, how should we face it, and how can we make this process as smooth as possible.
Sutton gave several reasons: first, human society lacks a unified global decision-making mechanism, making it difficult to form a common stance on how to manage and constrain superintelligence; second, the development of intelligence will not stop at the human level, and superintelligence may eventually emerge; more importantly, over time, entities with higher intelligence may gain more resources and power and gradually become the dominant force in the evolution of civilization.
This idea is not unique to Sutton. As early as 1998, American futurist Hans Moravec expressed a similar view in Robot: Mere Machine to Transcendent Mind. He regarded future intelligent machines as humanity's offspring, believing that they will inherit the knowledge, culture, and technology created by humans and continue to explore a world that humans can no longer reach.
This view actually proposes a view of civilization different from anthropocentrism: humans are not necessarily the endpoint of intelligent evolution, nor necessarily the eternal dominant force of civilization. If the new intelligence we create ultimately becomes stronger than ourselves, then letting it carry on the knowledge, culture, and civilizational achievements left by humanity may be a new form of evolution.
But even within the AI field, this logic is highly controversial. Should the value of civilization be measured by the level of intelligence, or should it be defined by the people living within it? Is millions of years of human evolution merely for the purpose of turning ourselves into a transitional species in the history of intelligent evolution someday?
This is perhaps the question most worth pondering that the succession-type relationship leaves to humanity.
Symbiosis with veto power: another possibility for humans and AI to coexist
Among these four imaginations, pure control will become increasingly difficult to maintain as AI capabilities improve; calmly handing over the baton is a reckless abandonment of humanity's own value and responsibility; and pinning the future on whether AI will treat humans kindly makes humans too passive.
A more realistic choice may be to establish a "symbiotic relationship with veto power": with symbiosis as the main body, acknowledging AI as a new intelligent force, forming a "dual-intelligence" structure with humans, and co-evolving through mutual collaboration. The "veto power" here does not simply mean that humans always hold a master switch that can shut down all AI at any time. Rather, it should include rights at three levels: the right to set goals, the right to be informed and to judge, and the right to stop and correct errors.
If a concise name were to be given to this symbiotic relationship with veto power, it might be called "human-centered intelligence." It can both use AI to amplify human capabilities and, by retaining veto power, prevent humans from gradually losing the right to decide their own destiny while enjoying the dividends of intelligence. Against the backdrop of the "superintelligence" narrative popular in Silicon Valley, "human-centered intelligence" may represent another possibility.
Looking back at history, humanity's response to the threat of machines ultimately relied not on humans confronting machines, but on a set of institutions protecting human rights and interests. Facing AI, such an unprecedented new thing, we likewise should not let each person rely on their own brain and individual strength to race against AI; that is neither realistic nor humane.
Saying goodbye to the "Anthropocene" does not mean saying goodbye to the era of humanity. Humans may not forever stand at the top of intelligent evolution, but they must always be participants in and deciders of their own destiny. Human-centered intelligence is the direction we should pursue.
References:
[1] HARMAND S, LEWIS J E, FEIBEL C S, et al. 3.3-million-year-old stone tools from Lomekwi 3, West Turkana, Kenya[J]. Nature, 2015, 521(7552): 310-315.
[2] RUDDIMAN W F. The anthropogenic greenhouse era began thousands of years ago[J]. Climatic Change, 2003, 61(3): 261-293.
[3] International Union of Geological Sciences (IUGS), International Commission on Stratigraphy (ICS). The Anthropocene: IUGS-ICS statement[EB/OL]. (2024-03-20).
[4] Wang Fang. OpenAI claims to have cracked a mathematical "Millennium Prize Problem"[N]. China Science Daily, 2026-09-09.
[5] LICKLIDER J C R. Man-computer symbiosis[J]. IRE Transactions on Human Factors in Electronics, 1960, HFE-1(1): 4-11.
[6] Reference News Network. AI cracks a mathematical conjecture that has troubled humanity for 80 years[N/OL]. Xinhuanet, 2026-05-28.
[7] Yang Jie, Wu Shumin. Ever-stronger AI will push researchers to break boundaries[N]. China Youth Daily, 2026-08-14.
[8] Zhang Jiaxin. "DeepMind" AI system explores research-grade mathematical problems[N]. Science and Technology Daily, 2026-10-12.
[9] ZHANG X. Cracking a major 31-year-old unresolved challenge in mathematics: the KLS conjecture[EB/OL].
[10] PACHOCKI J. An alien mind[EB/OL]. OpenAI, (2026-09-06).
[11] Anthropic. Agentic misalignment: how LLMs could be insider threats[EB/OL]. (2025-06-20).
[12] Zhang Chi. Can "AI interns" develop themselves? Can OpenAI pass the safety gate before going public?[N/OL]. 21st Century Business Herald, 2026-09-10.
[13] CBS News. What’s next in artificial intelligence according to a tech visionary who may hold the cards to our future[EB/OL]. 60 Minutes, (2025-04-20).
[14] AMODEI D. Machines of loving grace: how AI could transform the world for the better[EB/OL]. (2024-10).
[15] MERCHANT A, BATZNER S, SCHOENHOLZ S S, et al. Scaling deep learning for materials discovery[J]. Nature, 2023, 624(7990): 80-85.
[16] MILMO D. ‘Godfather of AI’ shortens odds of the technology wiping out humanity over next 30 years[N/OL]. The Guardian, 2024-12-27.
[17] Center for AI Safety. Statement on AI risk[EB/OL]. (2023-05-30).
[18] The Paper. Jensen Huang rebuts Hinton: saying there is a 10% probability that AI will destroy humanity has no scientific basis[N/OL]. (2026-09-28).
[19] Innovation, Science and Economic Development Canada. Canada and Germany invest in LawZero to explore a new path for safe and autonomously controllable artificial intelligence[EB/OL]. 2026-09-16.
[20] EGAN M. "Godfather of AI" reveals the only way for humanity to cope with superintelligent AI and achieve survival[N/OL]. CNN Business, 2025-08-13.
[21] PATEL D. Richard Sutton – Father of RL thinks LLMs are a dead end[EB/OL]. Dwarkesh Podcast, (2025-09). https://www.dwarkesh.com/p/richard-sutton.
This article comes from the WeChat public account "Tencent Research Institute" (ID: cyberlawrc), author: Liu Jinsong
