Peter Norvig: Artificial Intelligence, History and Future
The Evolution of Artificial Intelligence and Textbooks
Peter Norvig, Google's Director of Research and co-author of the seminal textbook Artificial Intelligence: A Modern Approach, discusses the transformation of AI over decades.
• Hardware Constraints: Early AI research was limited by memory and compute, shifting from predicate logic to first-order logic. The advent of GPUs and TPUs has enabled the deep learning revolution.
• Shifts in Methodology: The field has moved from symbolic AI and manual knowledge engineering toward probability, machine learning, and data-driven approaches.
• Defining Utility: Instead of just optimizing for a given utility function, researchers are increasingly grappling with the philosophical challenge of how to define those utility functions and human values.
Ethical AI and Societal Impact
Norvig addresses critical challenges in deploying AI systems in the real world:
• Fairness and Trade-offs: In areas like recidivism prediction, achieving fairness across protected classes often requires mathematical trade-offs between different definitions of equity.
• Adversarial Systems: We must move beyond low-dimensional metaphors for machine learning. AI models operate in high-dimensional spaces, making them vulnerable to adversarial attacks that are not immediately intuitive to humans.
• Marketplace of Attention: Digital platforms are incentivized to compete for user attention, which can conflict with long-term human well-being. He advocates for aligning incentives so systems work for users rather than just at them.
Education, Programming, and Future Outlook
"I think being one of the first classes, we were helped by sort of exterior motivation... But really a lot of it was, hey, this is a new thing and I'm really excited to be part of a new thing."
• MOOCs and Motivation: While online platforms provide access to education, community and personal motivation remain the primary drivers of success, not just the content itself.
• Programming as Problem Solving: Proficiency in programming is less about syntax and more about modeling, testing, and extracting knowledge from data.
• Future Goals: Norvig is excited about building AI assistants that can perform deep, real-world tasks and utilizing machine learning to assist developers in writing code by predicting potential bugs.