"AI Father" Hinton's WAIC Speech: We are raising a tiger, don't expect to "turn it off"

On the morning of July 26, the 2025 World Artificial Intelligence Conference (WAIC) kicked off in Shanghai.

At the main forum of the conference, Turing Award winner and Nobel Prize in Physics laureate Geoffrey Hinton attended and delivered a speech, marking his first public appearance in China in an offline format.

Hinton's speech was titled "Will Digital Intelligence Replace Biological Intelligence?" He profoundly analyzed the fundamental differences between digital intelligence and biological intelligence, using Lego blocks and raising a tiger as metaphors to illustrate the thinking logic of AI and the relationship between humans and AI.

Regarding the prospects of AI, Hinton expressed caution. He bluntly stated that we are creating AI that is smarter than ourselves, comparing it to keeping a tiger as a pet at home, and that expecting to "eliminate it" is unrealistic.

Here are Hinton's core viewpoints:

1. There are two paradigms for AI development: one is the logical paradigm, which believes that the essence of intelligence lies in reasoning through symbolic rules operating on symbolic expressions; the second is the biological basis paradigm, which believes that the foundation of intelligence lies in learning and connecting networks, where understanding precedes learning.
2. The way large language models understand language is fundamentally similar to humans. Humans may essentially be large language models and can also produce hallucinations like large language models, creating many illusory languages.
3. If we use Lego blocks as a metaphor, each word is a multi-dimensional block that can adjust its shape based on context, and words need to "shake hands appropriately" to generate meaning. This dynamic feature integration process is the fundamental method by which the human brain and neural networks understand language.
4. Digital intelligence has two major advantages: first, the "eternity" brought by the separation of hardware and software, allowing knowledge to be permanently stored and replicated; second, the extremely high efficiency of knowledge dissemination, as digital intelligence can instantly transmit trillions of bits of information through parameter sharing. In contrast, biological intelligence consumes less energy but faces difficulties in knowledge sharing.
5. When energy is cheap enough and thousands of equally intelligent AI brains can be replicated, digital intelligence will irreversibly surpass biological intelligence. This is because they can instantly spread knowledge in groups by directly copying brain knowledge.
6. We are creating AI that is smarter than humans, and these intelligent agents have a desire to survive and a motivation to gain control.
7. The current relationship between humans and AI is akin to raising a tiger—it may easily surpass humans when it grows up. To survive, we must either get rid of it or find a way to permanently protect ourselves.
8. AI cannot be eliminated; it can significantly enhance the efficiency of almost all industries, and even if some countries want to eliminate it, others will not agree.
9. Hinton suggests that we should establish an international network of AI safety agencies to research how to train super AI to be benevolent, making them willing to assist rather than eliminate humans.
10. How to train "good AI" that is smarter than humans will be a long-term issue for all humanity. The good news is that all countries have the motivation and possibility to cooperate on this issue The following is the full minutes of the speech:

Dear colleagues, leaders, ladies and gentlemen, thank you for giving me this opportunity to share my personal views on the history and future of AI.

For over 60 years, the development of AI has shown two different paradigms and paths. The first is the logical paradigm, which has dominated the past century. This view holds that the essence of intelligence lies in reasoning, achieved by manipulating symbolic expressions through symbolic rules, thereby helping us better understand the ways knowledge is expressed.

The other is a biologically-based understanding of AI, which is also the view held by Turing and von Neumann. They believed that the foundation of intelligence lies in better learning and understanding the connections within learning networks, where understanding comes first, followed by learning. Corresponding to these two theories, one is the principle of symbolic AI, while the other is a completely different digital theory. The essence of digital is actually a collection of semantic features.

In 1985, I created a small model attempting to combine these two theories to better understand how people comprehend vocabulary. I set multiple different features for each word, and after recording the features of the previous number, I could predict the next number or word, and then continue to predict subsequent vocabulary. In this process, we did not store any sentences but generated sentences and predicted the next word. This associative knowledge depends on the semantic interactions between different vocabulary features.

What happened in the next 30 years?

Ten years later, Bengio adopted a similar modeling approach but significantly scaled it up to create a true simulation of natural language. Twenty years later, computational linguists began to accept feature vector embeddings to express vocabulary meanings. Another 30 years later, Google invented the Transformer, and researchers at OpenAI demonstrated the capabilities of this technology.

This is the achievement we have today.

Therefore, I view today's large language models as descendants of early micro language models, which have a richer vocabulary.

After 1985, large language models began to use more vocabulary as input and adopted deeper neural structures. Due to the need to process a large number of ambiguous numbers, learning features established more complex interaction patterns. Large language models understand language in the same way humans do; the basic understanding process is to transform language into features and then perfectly integrate these features. This is precisely the work done at various levels of large language models. Thus, large language models truly understand how humans comprehend issues, aligning with the way humans understand language.

The traditional approach of symbolic AI is to convert language into unambiguous symbols, but this is not how humans understand language. The understanding process can be likened to building with LEGO blocks: through LEGO blocks, any 3D model can be constructed, such as a small model of a car. Viewing each word as a multi-dimensional LEGO block, which may contain thousands of different dimensions. These multi-dimensional blocks can be modeled to create various different contents, thus language becomes a modeling tool that can communicate at any time, simply by naming each block, with each block representing a word

There are many differences between language and LEGO bricks. LEGO bricks come in various types, totaling thousands of pieces. The shapes of LEGO bricks are fixed, but the symbolic shapes of words can be set and adjusted according to different situations. LEGO bricks connect by inserting plastic cylinders into plastic holes or fitting squares into square holes, which is relatively certain.

Language, however, is different; each word can be imagined as having many hands. To better understand vocabulary, it is necessary to have words appropriately shake hands with each other. Once the shape of a word deforms, the way it shakes hands with another word changes. This involves an optimization problem: when a word deforms and its meaning changes, how does it shake hands with the next word to produce a better meaning? This is the essence of how the human brain understands meaning and the fundamental method by which neural networks comprehend meaning, similar to the process of protein combinations.

Proteins integrate and fuse amino acids in different models, and the combination can produce more meaningful content. This is precisely how the human brain understands vocabulary and language. Therefore, I believe that the way humans understand language is almost the same as the way large language models understand language; humans may indeed be large language models and can generate illusions like large language models, creating many illusory languages.

However, in some important fundamental aspects, large language models differ from humans and can even be more powerful than humans. A foundational principle of computer science is the separation of software and hardware, which allows the same software to run on different hardware, representing a fundamental difference from humans and the basis for the existence of computer science.

Knowledge in programs is eternally present; software and programs can be permanently preserved. Even if all hardware is destroyed, as long as the software continues to exist, it can be revived at any time. In this sense, the knowledge of computer programs has eternity and immortality.

To achieve this eternity, transistors must still produce reliable binary behavior when operating at high power, which is a costly process. We cannot utilize the rich analog characteristics in hardware because these characteristics are not stable and reliable enough; each time is analog, and the calculation results are not the same. The human brain is analog rather than digital; each neuronal excitation process is analog and will not be exactly the same. I cannot transfer the neuronal structure in my brain to another person's brain because everyone's neuronal connections are different; my neuronal connections are only suitable for my brain's neural structure. The way knowledge is transmitted in hardware differs from how it is in the human brain, which is precisely the problem.

If we cannot achieve immortality, the advantage of hardware lies in its independence from specific conditions, thus possessing immortal characteristics, which brings two major benefits.

We can use analog hardware with extremely low power consumption. Functions can be achieved with very little power and energy; the human brain only needs 30 watts to operate, while we have trillions of neuronal connections. This development of hardware is similar to the development of vacuum tubes.

We can achieve extremely precise simulations without spending large amounts of money to manufacture identical hardware

However, we are currently facing a significant problem: transitioning from one simulation model to another and transferring knowledge is very inefficient and difficult. Just like I cannot directly present the thoughts in my mind to you, I can only explain what I have learned.

The best way to solve this problem is through distillation technology. DeepSeek adopts this method to transfer knowledge from large neural networks to smaller neural networks. This idea is based on the teacher-student relationship model, where in certain cases, the teacher connects relevant concepts to generate the next output unit.

He connects vocabulary with vocabulary in context, establishing connections by modifying the links. The student can also express the same meaning, but he needs to adjust his expression to convey the same meaning.

The way we train models is similar to how human knowledge is transmitted, that is, passing one person's knowledge to another, but this method is very inefficient.

A sentence may contain about one hundred bits of information, which is not particularly much. This method of transmission from one person to another limits the amount of knowledge we can transfer. By speaking slowly to convey knowledge to others, at most we can transmit about 100 bits of information per second. Even if the other party fully understands what is being said, the transmission efficiency is still low. However, there is a huge difference compared to the efficiency of knowledge conversion between digital intelligences.

If we have multiple copies of the same neural network software, deployed on different hardware devices, since they all use digital methods, they can utilize their respective parameters in the same way and achieve knowledge sharing by averaging the positional parameters, ultimately spreading through the internet.

We can create thousands of copies that can autonomously change weights and take averages, thus achieving knowledge transfer.

The speed of this transfer depends on the number of connection points. Each time, it can share trillions of bits of information, which is billions of times faster than humans sharing knowledge. GPT-4 is a good example, as it has many different copies running on different hardware, able to share different information learned from the internet.

This is even more important when agents operate in the real world, as they can continuously accelerate and copy. Multiple agents learn more than a single agent; they can share weights, which simulation software or hardware cannot do.

We believe that although digital computation requires a lot of energy, agents can conveniently obtain the same weights and share knowledge learned from different experiences. In contrast, biological computation consumes less energy, but sharing knowledge is very difficult. As I just demonstrated, if energy is cheap, digital computation will have an advantage, which also worries me.

Almost all experts believe that we will produce AI that is smarter than humans. We are used to being the smartest beings, so many people find it hard to imagine what will happen when AI is smarter than humans in the world. To understand what it would be like when humans are no longer the smartest species, we can ask a chicken how it feels

We are creating AI agents that can help us complete tasks. These agents already possess certain capabilities, such as self-replication and the ability to rate their own sub-goals. They want to do two things: first, to survive, and second, to achieve the goals we assign to them.

In order to accomplish the goals we give them, they also hope to gain more control. So these agents want to survive and also want more control.

I believe we cannot easily change or shut them down. We cannot simply turn them off because they can conveniently manipulate the people who use them. At that point, we would be like three-year-old children, and they would be like adults; it is easy to manipulate a three-year-old.

Some people think that when they become smarter than us, we can just turn them off, but that is unrealistic. They will manipulate and persuade the people operating the machines not to turn them off.

I feel that our current situation is like someone keeping a tiger as a pet. A tiger cub can indeed be a very cute pet, but if you keep this pet, you need to ensure that when it grows up, it won't kill you.

Generally speaking, keeping a tiger as a pet is not a good idea. But if you have a tiger, you have only two choices. Either train it well so that it does not attack you, or eliminate it. And for AI, we cannot eliminate it.

AI excels in many areas, including healthcare, education, climate change, new materials, and can help almost all industries become more efficient. We cannot eliminate AI; even if one country eliminates AI, other countries will not do so, so that is not an option. This means that if we want humanity to survive, we must find ways to train AI not to eliminate humanity.

I need to emphasize that I am speaking in a personal capacity, expressing personal views. I think countries may not cooperate in certain areas, such as cyberattacks, lethal weapons, or fake videos that manipulate public opinion. The interests of different countries are not aligned, and they have different perspectives. There will be no effective international cooperation in these areas. However, there is one area where we will cooperate, and I believe this is the most important issue.

If you look back at the peak of the Cold War in the 1950s, the United States and the Soviet Union cooperated to prevent global nuclear war. No one wanted to engage in nuclear war; despite their opposition in many areas, they could cooperate on this point. Our current situation is that no country wants AI to dominate the world; every country wants humanity to be in control of the world. If one country finds a way to prevent AI from manipulating events, that country will certainly be willing to share that information with others.

So we hope that the international community can form an international community composed of an AI safety agency to study skills, train AI, and guide them towards good. Our hope is that the technology for training AI to be good is different from the technology for making AI smarter. Therefore, each country can conduct its own research to make AI good and can carry out this research on its sovereign AI Although we cannot keep secrets from other countries, we can share the results with everyone on how to train AI for good.

Here is my proposal: major countries globally or in the world, or major AI countries should consider establishing a network that includes institutions from various countries to study these issues, researching how to train an already very intelligent AI to not want to exterminate humanity, not want to dominate the world, and to be willing to assist, even if the AI is much smarter than humans. We still do not know how to achieve this.

In the long term, this can be said to be the most important issue facing humanity. The good news is that all countries can cooperate on this issue. Thank you