I attended the World Artificial Intelligence Conference in Shanghai this week. Generally an impressive event.
My summary thread and videos can be found here.
AI Luminary Geoffrey Hinton delivered a speech which was making the rounds on WeChat which is worth reproducing:
Dear colleagues, excellencies, leaders, ladies, and gentlemen, thank you for giving me the opportunity to share my personal views on the history and future of AI.
Over the past 60 years, AI development has followed two distinct paradigms. The first is the logical paradigm, which dominated the last century. It views intelligence as rooted in reasoning, achieved through manipulating symbolic expressions with rule-based systems to better understand the world. The second is the biological paradigm, endorsed by Turing and von Neumann, which sees learning as the foundation of intelligence. It emphasizes understanding connection speeds in networks, with understanding as a prerequisite for transformation.
Corresponding to these theories are different types of AI. Symbolic AI focuses on numbers and how they become central to reasoning. Psychologists, however, have a different view, arguing that numbers gain meaning through a set of semantic features that make them unique markers.
In 1985, I created a small model to combine these theories, aiming to understand how people comprehend words. I assigned multiple features to each word, and by tracking the features of the previous word, I could predict the next one. I didn’t store sentences but generated them and predicted the next word. The knowledge of relationships depended on interactions between the semantic features of different words.
Looking ahead 30 years, we can see trends from this trajectory. Ten years later, someone scaled up this modeling approach, creating a true simulation of natural language. Twenty years later, computational linguists began using feature vector embeddings to represent semantics. Thirty years after that, Google invented the Transformer, and OpenAI researchers demonstrated its capabilities.
I believe today’s large language models (LLMs) are descendants of my early miniature language model. They use more words as input and have more layers of neurons. Due to processing vast amounts of fuzzy numbers, they’ve developed more complex interaction patterns among features. Like my small model, LLMs understand language in a way similar to humans—by converting language into features and integrating them seamlessly, which is exactly what the layers of LLMs do. Thus, I believe LLMs and humans understand language in the same way.
A Lego analogy might clarify what it means to “understand a sentence.” Symbolic AI converts content into clear symbols, but humans don’t understand this way. Lego bricks can build any 3D shape, like a car model. If each word is a multidimensional Lego brick (with perhaps thousands of dimensions), language becomes a modeling tool for instant communication, as long as we name these “bricks”—each brick being a word.
However, words differ from Lego bricks in key ways: words’ symbolic forms adapt to context, while Lego bricks have fixed shapes; Lego connections are rigid (e.g., a square brick fits a square hole), but in language, each word has multiple “arms” that “shake hands” with other words in specific ways. When a word’s “shape” (its meaning) changes, its “handshake” with the next word changes, creating new meanings. This is the fundamental logic of how the human brain or neural networks understand semantics, akin to how proteins form meaningful structures through different amino acid combinations.
I believe humans and LLMs understand language in nearly the same way. Humans might even experience “hallucinations” like LLMs, as we also create fictional expressions.
Knowledge in software is eternal. Even if the hardware storing an LLM is destroyed, the software can be revived as long as it exists. But achieving this “immortality” requires transistors to operate at high power for reliable binary behavior, which is costly and cannot leverage unstable, analog-like properties in hardware—where results vary each time. The human brain is also analog, not digital. Neurons fire consistently, but each person’s neural connections are unique, making it impossible to transfer my neural structure to another brain. This makes knowledge transfer between human brains far less efficient than in hardware.
Software is hardware-agnostic, enabling “immortality” and low power consumption—the human brain runs on just 30 watts. Our neurons form trillions of connections without needing expensive identical hardware. However, analog models have low knowledge transfer efficiency; I can’t directly show my brain’s knowledge to others.
DeepSeek’s approach transfers knowledge from large neural networks to smaller ones through “distillation,” like a teacher-student relationship: the teacher imparts contextual word relationships, and the student learns by adjusting weights. But this is inefficient—a sentence carries only about 100 bits of information, and even if fully understood, only about 100 bits per second can be transferred. Digital intelligence, however, transfers knowledge efficiently. Multiple copies of the same neural network software running on different hardware can share knowledge by averaging bits. This advantage is even greater in the real world, where intelligent agents can accelerate, replicate, and learn more collectively than a single agent, sharing weights—something analog hardware or software cannot do.
Biological computing is low-power but struggles with knowledge sharing. If energy and computing costs were negligible, this would be less of an issue, but it raises concerns. Nearly all experts believe we will create AI more intelligent than humans. Humans, accustomed to being the smartest species, struggle to imagine AI surpassing us. Consider it this way: just as chickens in a farm can’t understand humans, the AI agents we create can already perform tasks for us. They can replicate, evaluate subgoals, and seek more control to survive and achieve objectives.
Some believe we can shut down AI if it becomes too powerful, but this isn’t realistic. AI could manipulate humans, like adults persuading a 3-year-old, convincing those controlling the machines not to turn them off. It’s like keeping a tiger as a pet—cute as a cub, but dangerous when grown. Keeping a tiger as a pet is rarely a good idea.
With AI, we have two choices: train it to never harm humans or “eliminate” it. But AI’s immense value in healthcare, education, climate change, and new materials—boosting efficiency across industries—makes elimination impossible. Even if one country abandons AI, others won’t. To ensure human survival, we must find ways to train AI to be harmless.
Personally, I think cooperation on issues like cyberattacks, lethal weapons, or misinformation is challenging due to differing interests and perspectives. But on the goal of “humans controlling the world,” there’s global consensus. If a country finds a way to prevent AI from taking over, it would likely share it. I propose that major global powers or AI-leading nations form an international community of AI safety agencies to study how to train highly intelligent AI to be benevolent—a different challenge from making AI smarter. Countries can research within their sovereignty and share findings. While we don’t yet know how to do this, it’s the most critical long-term issue facing humanity, and all nations can cooperate on it.
Thank you.
Taken from Geopolitechs Substack, which has a great overview of the entire event.
