In Lecture 1 of Stanford CS221, Professor Percy Liang introduces the foundational concepts of AI—perception, reasoning, action, and learning—while discussing the evolution of the field through symbolic, neural, and statistical paradigms, and highlighting the societal and ethical implications of AI systems. The lecture concludes with a practical overview of tensors and their central role in modern machine learning, setting the stage for the course’s hands-on, code-driven approach.
In the first lecture of Stanford’s CS221: Artificial Intelligence Principles and Techniques, Professor Percy Liang introduces the course by reflecting on the rapid evolution of AI and the foundational concepts that remain central to the field. He begins by defining AI as the study of building artificial agents—systems that perceive their environment, reason about it, act upon it, and learn from experience. These four pillars—perception, reasoning, action, and learning—are explored through examples such as autonomous vehicles and language models, emphasizing that intelligence is not just a human trait but a general property that can be instantiated in machines.
Liang highlights the importance of resource constraints in AI, noting that agents must operate with limited computational power and incomplete information. This reality shapes the need for efficient algorithms and clever engineering. He also discusses the dual perspectives on AI objectives: the developer’s goals (such as alignment between intended and actual agent behavior) and broader societal impacts, including privacy, intellectual property, job displacement, and ethical considerations. The lecture stresses that AI systems inevitably encode values—explicitly or implicitly—and that aligning these with societal good is a complex, ongoing challenge.
The lecture then transitions to a historical overview of AI, tracing its development through three main paradigms: symbolic AI, neural AI, and statistical AI. Symbolic AI, rooted in logic and rule-based systems, saw early optimism but faced setbacks due to computational limitations and the complexity of encoding real-world knowledge. Neural AI, inspired by biological neurons, experienced cycles of enthusiasm and skepticism, eventually gaining prominence with the advent of deep learning and breakthroughs in areas like computer vision and language modeling. Statistical AI, grounded in mathematical and probabilistic methods, contributed foundational techniques such as linear regression and Markov decision processes, many of which predate the modern AI era.
Liang notes the recent emergence of foundation models—large-scale, pre-trained neural networks that have transformed natural language processing and other domains. He observes that the field has shifted from a niche academic pursuit to a major industrial and societal force, with massive investments and increasing secrecy around model development. Despite remarkable progress, many fundamental questions about intelligence and reasoning remain open, and the interplay between symbolic, neural, and statistical approaches continues to shape the field.
The lecture concludes with a practical introduction to tensors, the core data structure in modern machine learning. Liang explains how tensors generalize scalars, vectors, and matrices to higher dimensions, and demonstrates their use in representing data, model parameters, and computations. He introduces tools like NumPy and the einops library for efficient and readable tensor manipulation, emphasizing the importance of expressing computations in terms of tensor operations for both performance and clarity. The hands-on, code-driven approach of the course is highlighted, preparing students to engage deeply with the empirical and engineering aspects of AI.