CS50x 2026 – Artificial Intelligence

Harvard's Introduction to Artificial Intelligence

CS50x 2026 presents Harvard University's comprehensive introduction to artificial intelligence, delivered by renowned instructor David J. Malan. This 48-minute lecture serves as the foundational entry point into the intellectual enterprises of AI, bridging theoretical concepts with practical applications. The course tackles the essential components of modern AI systems, from classical machine learning approaches to cutting-edge generative models that have transformed how computers process information and generate human-like responses.

Generative AI and Its Impact

The lecture opens by addressing generative artificial intelligence, one of the most transformative developments in computer science. Generative AI systems are designed to create new content—text, images, code, and more—based on patterns learned from vast datasets. This technology underpins tools like ChatGPT, Claude, and other large language models that have captured mainstream attention. Understanding how these systems work requires grasping both the mathematical foundations and the practical limitations that arise when training models on real-world data.

Mastering Prompt Engineering

Prompt engineering emerges as a crucial skill in the AI era. Rather than simply asking an AI system a question, prompt engineering involves crafting carefully structured requests that guide generative models toward desired outputs. This discipline bridges the gap between user intent and model capability, enabling developers and end-users to extract maximum value from language models. The lecture explores how phrasing, context, and instruction clarity dramatically influence response quality, making prompt engineering both an art and a science.

Practical Tools: CS50.ai and Copilot

The course introduces CS50.ai, a tool developed for the CS50 community, and Microsoft's Copilot, demonstrating how AI assistants integrate into real workflows. These tools exemplify the practical deployment of large language models, showing students how theoretical AI concepts manifest in products used by millions. By examining these systems, learners gain insight into how prompt engineering translates into user-facing features and how different AI implementations handle similar tasks.

Classical Machine Learning Foundations

Before diving into modern deep learning, the lecture establishes classical machine learning concepts. Decision trees are presented as an intuitive machine learning algorithm that mirrors human decision-making processes. Decision trees partition feature space recursively, enabling the model to learn non-linear relationships in data. The minimax algorithm is then introduced as a game-theoretic approach to decision-making, illustrating how machines can think strategically in adversarial scenarios. These classical methods provide essential groundwork for understanding how machines learn patterns from examples.

The Shift to Deep Learning

The course transitions to machine learning more broadly, exploring how systems learn directly from data rather than relying on explicit programming. Machine learning encompasses supervised learning, unsupervised learning, and reinforcement learning paradigms. Deep learning represents a significant evolution, using artificial neural networks with multiple layers to learn hierarchical representations of data. These deep neural networks have revolutionized computer vision, natural language processing, and many other domains by discovering features automatically rather than requiring manual feature engineering.

Large Language Models and Modern AI

Large language models (LLMs) represent the pinnacle of current generative AI research. These models are trained on billions of text examples, learning statistical patterns that enable them to predict the next token in a sequence with remarkable accuracy. The scale of these models—measured in billions or trillions of parameters—combined with transformer-based architectures enables emergent abilities like few-shot learning, reasoning, and code generation. Understanding LLMs requires grasping both the transformer architecture's attention mechanisms and the practical implications of training and deploying models of unprecedented scale.

Addressing AI Hallucinations

The lecture concludes by confronting a critical limitation of generative AI: hallucinations. Language models, despite their sophistication, sometimes generate plausible-sounding but factually incorrect information. This phenomenon arises because LLMs predict text based on statistical likelihood rather than verifiable truth. Hallucinations represent a fundamental challenge in deploying AI systems for applications requiring high factual accuracy, such as medical diagnosis, legal documentation, or scientific research. Recognizing and mitigating hallucinations is essential for responsible AI deployment.

Bridging Theory and Practice

Throughout this lecture, CS50 maintains its tradition of connecting abstract computer science concepts to real-world applications. The progression from decision trees to deep learning to large language models shows how AI capabilities have evolved, and how each technique remains relevant in modern systems. Students emerge with both the conceptual frameworks necessary to understand AI research papers and practical knowledge for applying these tools in projects, whether building chatbots, training custom models, or prompting existing systems effectively.