Transcript of Geoffrey Hinton's 2025 WAIC speech

Colleagues, Excellencies, Leaders, Ladies and Gentlemen, first of all, thank you very much for giving me this opportunity to share my personal perspective on the history and future of AI.

For over 60 years, there have been two distinct paradigms and paths in the development of AI. One is the logical paradigm, which has been the mainstream for the past century. It holds that the essence of intelligence lies in reasoning, achieved by manipulating symbolic expressions through symbolic rules, thereby helping us better understand the world. The other is the biologically based paradigm, endorsed by Turing and von Neumann. They believe that the foundation of intelligence is learning and understanding the speed of connections in a network. Understanding is a prerequisite for transformation.

Corresponding to these two theories are different types of AI. Symbolic AI focuses on numbers, but psychologists have a completely different theory about how these numbers become central. They believe that the meaning of numbers lies in a series of semantic features that make them unique symbols.

In 1985, I developed a small model that attempted to combine these two theories to understand how people understand words. I assigned multiple different features to each word. By recording the features of the previous word, I could predict the next word. In this process, I didn't store any sentences; I generated sentences and predicted the next word. The knowledge of correlations depends on how the semantic features of different words interact.

If you ask what will happen in the next 30 years, the trajectory of development reveals some trends. Ten years later, some people continued this modeling approach, but significantly scaled it up, making it a true simulation of natural language. Twenty years later, computational linguists began to embrace feature vector embeddings to represent semantics. Another 30 years later, Google invented the Transformer, and researchers at OpenAI demonstrated its capabilities.

So, I think of today's large language models as the "descendants" of my tiny language models. They use more words as input and employ more layers of neurons. Because they have to process a large number of fuzzy numbers, they also learn more complex interactions between features. But like my small model, large language models understand language similarly to humans—the fundamental logic is to convert language into features and then integrate these features in a seamless manner. This is exactly what each layer of a large language model does. Therefore, I believe that large language models understand language in the same way that humans do.

The analogy of Lego bricks might better explain what it means to "understand a sentence." Symbolic AI transforms content into clear symbols, but humans don't understand it that way. Lego bricks can be assembled into any 3D shape, such as a car model. If we think of each word as a multi-dimensional Lego brick (perhaps thousands of dimensions), language becomes a modeling tool that can be used to communicate with people at any time. All we need to do is give these "bricks" a name—each "brick" is a word.

However, words differ from Lego bricks in many ways: the symbolic form of a word can be adjusted according to the situation, while Lego bricks have a fixed shape; Lego bricks are fixed in their connection (for example, a square brick fits into a square hole), but in language, each word is like multiple "arms" that interact with other words through appropriate "handshakes." As the word's "shape" changes, so does the "handshake." When the "shape" (i.e., meaning) of a word changes, the way it "handshakes" with the next word changes, resulting in new meaning. This is the fundamental logic behind how the human brain or neural network understands semantics, similar to how proteins form meaningful structures through different combinations of amino acids.

Thus, I believe that humans understand language in much the same way as large language models. We may even experience "hallucinations" like large language models, as we also create fictional expressions.

Knowledge in software is immortal. Even if the hardware storing the LLM is destroyed, as long as the software exists, it can be "resurrected" at any time. However, to achieve this "immortality," transistors must operate at high power to produce reliable binary behavior. This process is costly and cannot take advantage of the similarly unstable characteristics of hardware—it is analog, and the results of each calculation are different. The human brain is also analog, not digital. The process of neurons firing is the same every time, but each person's neurons are connected differently. I cannot transfer my neural structure to another person's brain. As a result, the efficiency of knowledge transmission between people is far lower than that of knowledge transmission within hardware.

Software is hardware-independent, making it "immortal" and offering the advantage of low power consumption—the human brain only requires 30 watts to operate. Our neurons have trillions of connections, eliminating the need for the massive expense of manufacturing identical hardware. However, the problem is that knowledge transfer between simulation models is extremely inefficient, making it impossible to directly share knowledge from my brain with others.

Deepseek's approach involves transferring knowledge from large neural networks to smaller ones, a process known as "distillation," similar to the relationship between a teacher and a student: the teacher teaches the contextual associations of words to the student, who then learns to express them by adjusting weights. However, this approach is inefficient. A sentence typically contains only 100 bits of information, and even if fully understood, only about 100 bits can be transferred per second. Knowledge transfer between digital intelligences, however, is extremely efficient. When multiple copies of the same neural network software run on different hardware, they can share knowledge by averaging bits. This advantage is even more pronounced when intelligent agents operate in the real world—they can continuously accelerate and replicate, allowing multiple agents to learn more than a single one and share weights—something that cannot be achieved with analog hardware or software.

Biological computing offers low power consumption, but knowledge sharing is difficult. If energy and computing costs were low, the situation would be much better, but this also worries me—almost all experts agree that we will create AI that is more intelligent than humans. Humans are so accustomed to being the most intelligent creatures that it's hard to imagine AI surpassing us. But here's another way to look at it: just as chickens on a farm can't understand humans, the AI agents we've created can already help us complete tasks. They can copy themselves, evaluate sub-goals, and seek more control to survive and achieve their goals.

Some believe that AI can be shut down when it becomes too powerful, but this is unrealistic. They might manipulate humans like an adult manipulating a three-year-old, convincing those controlling the machines not to shut them down. It's like keeping a tiger as a pet: a cub is cute, but it could become harmful when it grows up, and keeping a tiger as a pet is generally not a good idea. With AI, we have two choices: either train it to never harm humans, or eliminate it. However, AI has enormous potential in areas like healthcare, education, climate change, and new materials, and it can improve efficiency across all industries. We can't eliminate it—even if one country abandons AI, others won't. Therefore, if humanity is to survive, we must find ways to train AI to not harm humans.

I personally believe that cooperation among countries in areas such as cyberattacks, lethal weapons, and disinformation manipulation is difficult due to differing interests and perspectives. However, on the goal of "human control of the world," all countries agree: if a country finds a way to prevent AI from dominating the world, it will certainly be willing to share. Therefore, I propose that major countries or AI-powered nations establish an international community composed of AI safety institutions to research how to train highly intelligent AI for good—this is different from training AI to be smart. Countries can conduct research within their own sovereign domains and then share the results. While the specifics of how this will be done are still unclear, this is one of the most important long-term issues facing humanity, and all countries can cooperate in this area.

Thank you.