Summary of Key Points
At the 2026 World Artificial Intelligence Conference (WAIC), Richard Sutton, known as the "father of reinforcement learning" and the recipient of the 2024 Turing Award, shared his entrepreneurial philosophy, the future direction of AI technology (shifting from reliance on human data to "experience-based learning"), the practice of embodied intelligence (through the Robot Kindergarten project), and advice for young people's development. He emphasized the complementary role of for-profit and non-profit organizations in supporting scientific research, stating that AI must return to its fundamental purpose of learning from experience. He also suggested that robots should develop through trial and error, just like babies, and that young people should cultivate independent thinking.
I. Entrepreneurial Philosophy: The "Dual Drive" of For-Profit and Non-Profit
Sutton founded the for-profit company Oak Lab not solely for profit but to complement his non-profit organization Openmind:
- Role of Non-Profits: Openmind focuses on making knowledge open-source and disseminating principles of intelligence, ensuring that AI is accessible to everyone while reducing the risk of technology being misused (e.g., preventing a few individuals from monopolizing core technologies).
- Value of For-Profit Organizations: Cutting-edge AI research requires significant funding (e.g., purchasing computing power and hiring teams). For-profit companies can generate revenue through short-term patent sales, which in turn supports further research. Although technology will eventually be made open-source, temporary protection helps sustain innovation.
In summary, non-profits are responsible for ensuring accessibility and security, while for-profit organizations provide the financial resources needed for continuous development; both are indispensable.
II. The "Big Shift" in AI Technology: Moving Away from Human Data towards "Experience-Based Learning"
Current large language models (such as ChatGPT) are trained using vast amounts of human data (articles, videos, conversations), but Sutton believes this approach is misplaced:
- The Problem: Traditional models are energy-intensive (e.g., consuming millions of watts during training), whereas the human brain can handle complex tasks with just 60 watts. This is due to the von Neumann architecture used in computers, which requires data to be constantly transferred between storage and processing units, leading to inefficiencies.
- Future Direction: AI should adopt a parallel architecture similar to the human brain, allowing data to remain in place without unnecessary transfers, thereby significantly improving efficiency. More importantly, AI should learn from experience, just like AlphaGo—by making mistakes and adjusting based on its own experiences (e.g., a robot learning to walk by trial and error). Sutton argues that the industry has been overly reliant on human data for too long and that shifting to experience-based learning will become the mainstream in the next decade.
III. Breakthroughs in Embodied Intelligence: Letting Robots Learn through Trial and Error
Sutton's Robot Kindergarten project addresses a major limitation of traditional robots:
- Traditional Robots: They can only perform tasks based on instructions and often fail when they make mistakes (e.g., breaking down after a fall), lacking the ability to learn skills through trial and error like babies.
- How the Project Works: The goal is to create robots that can withstand failures and learn through repeated attempts in a dedicated environment. For example, a robot might try to pick up a cup, adjust its grip if it falls, and continue until it succeeds. This approach helps robots acquire basic skills.
- Future Potential: Once robots master these skills, they could assist with tasks like caring for the elderly or teaching children, potentially filling labor gaps (e.g., due to an aging population).
IV. Advice for Young People: Independent Thinking is More Important than Chasing Trends
Sutton offers practical advice to two groups of young people:
1. AI Researchers: Don't blindly follow authorities; use fundamental principles to question the nature of AI and develop your own ideas. Writing a page of notes daily can help clarify your thoughts, and persistence will yield results.
2. General Young People: Distinguish between "computing" (a tool, like using a computer for work) and "intelligence" (the ability to think creatively). Generalized intelligence could emerge in the next 14 years, so build your own cognitive foundations by delving deeply into a field and learning to analyze problems independently. Remember: there are no absolute authorities in science; be responsible for your own ideas and verify them with facts.
V. The AI Trends of the Next Decade: The "Age of Experience"
Sutton predicts that experience-based learning will become the core focus of the AI industry over the next decade:
- Investors are already investing in this area (e.g., Yang Likun's company is doing so), and Oak Lab has successfully raised funds.
- In 8 years, experience-based learning will be as crucial as current large language models. By then, AI will no longer rely on human data but will learn and evolve independently in real-world scenarios, becoming more intelligent and efficient.
In summary, Sutton's main message is that AI should return to its essence—learning from experience rather than relying on human data. Scientific research requires the collaboration of for-profit and non-profit organizations to achieve long-term progress. For young people, independent thinking and expertise in a specific field are the best preparations for the changes brought about by AI.