Stanford CME295 Transformers & LLMs | Autumn 2025 | Lecture 3 - Tranformers & Large Language Models

This lecture explains Artificial Intelligence (AI) in clear, simple terms and shows how it fits into everyday life. It begins with a classic definition from Marvin Minsky: AI is the science of making machines do things that would need intelligence if a person did them. The talk then divides AI into two main branches. Narrow (or weak) AI focuses on specific tasks—like recognizing faces, playing chess, or recommending products—and is the kind we use most today. General (or strong) AI is a hypothetical future system that could match a human’s ability to learn and reason across any topic, but it does not exist yet. After setting this foundation, the lecture moves into the main ways we actually build AI systems: machine learning, natural language processing (NLP), and computer vision.

The lecture gives special attention to machine learning because it powers much of modern AI. Instead of writing step-by-step rules for every possible situation, we give the computer data, and it learns patterns from that data. The lecture introduces three common types of machine learning. Supervised learning uses labeled examples (like images marked “cat” or “dog”) to teach a model to predict correct labels for new examples. Unsupervised learning works without labels and tries to find hidden structure, such as grouping customers into segments by behavior. Reinforcement learning teaches agents by giving them rewards or penalties for their actions, helping them learn strategies over time, which is how many systems learned to play games like chess and Go.

Next, the lecture explains two important application areas: NLP and computer vision. NLP lets computers understand, interpret, and generate human language. It enables tasks like machine translation, chatbots, and answering questions. Computer vision allows machines to “see” and make sense of images and video. It supports object detection, face recognition, and motion tracking, which power self-driving cars, security systems, and medical image analysis.

To connect these ideas to the real world, the lecture surveys where AI is used today: healthcare (diagnosing diseases and suggesting treatments), finance (fraud detection and risk management), transportation (self-driving cars and traffic optimization), entertainment (recommendation engines for movies and music), manufacturing (automation and quality control), and retail (personalized shopping and supply chain optimization). The main message is that AI is already deeply woven into many industries and everyday tools, and its impact is growing.

The lecture also addresses risks and ethics. It highlights three core concerns: job displacement as AI automates certain tasks; bias in AI systems when training data is skewed or unfair, leading to harmful outcomes; and misuse of AI for malicious ends, such as autonomous weapons or invasive surveillance. The lecture emphasizes the need for safeguards and ethical guidelines to ensure AI is developed and used to benefit people while minimizing harm. The overall structure moves from defining AI and its types, to core methods (especially machine learning and its three main styles), to key application domains, and then to societal impacts and responsibilities. By the end, you understand what AI is, how it is built, where it is used, what can go wrong, and how to think about building and using it responsibly.

This lecture is ideal for beginners who want a broad but solid introduction. You do not need math or programming to follow it, though some familiarity with everyday apps (like recommendation systems or translation tools) helps. After studying this material, you will be able to explain narrow vs. general AI, describe the main kinds of machine learning with examples, name what NLP and computer vision do, recognize common uses of AI across industries, and discuss key risks and ethical needs. The content is organized to give a complete picture, so you can confidently talk about AI in school, at work, or in everyday conversations.

Overall Architecture/Structure

To understand modern AI in practice, imagine a simple pipeline that most AI projects follow, even if the exact tools change. First, define the problem: what is the task, who is affected, and how will success be measured? Second, gather data: this may be images, text, audio, or tabular records. Third, prepare the data: clean errors, balance classes, and split into training, validation, and test sets. Fourth, choose an approach that matches the problem: supervised learning for labeled prediction tasks, unsupervised learning for structure discovery, or reinforcement learning (RL) for sequential decisions. Fifth, train the model: the system adjusts its internal settings (parameters) to reduce mistakes on the training data. Sixth, evaluate and iterate: measure performance, check for bias, and tune until the model generalizes well to new data. Finally, deploy and monitor: integrate the model into an app or workflow and watch it in the real world, ready to update it when data or goals change.

Role of Each Component

• Problem definition sets scope and prevents solving the wrong problem. Clear objectives (e.g., “identify cats vs. dogs in photos with at least 95% accuracy”) guide every later step.
• Data collection brings the raw material the model learns from. The quality, diversity, and representativeness of data strongly affect fairness and accuracy.
• Data preparation ensures inputs are consistent and useful. For images, this might mean resizing and normalizing; for text, tokenizing; for tables, handling missing values and scaling features.
• Model selection matches task structure: supervised for predicting labels, unsupervised for finding groups, RL for learning by reward. NLP and vision tasks often use specialized models suited to sequences or images.
• Training adjusts parameters to minimize error. In supervised learning, this usually means comparing predictions to labels and updating the model to reduce the difference. In unsupervised learning, the model seeks compact clusters or low-dimensional patterns. In RL, the agent explores actions and updates a policy to maximize long-term reward.
• Evaluation checks accuracy and other metrics, but also fairness, robustness, and safety. A model that is accurate on average but biased for a subgroup is not acceptable.
• Deployment integrates the model into products, with monitoring for drift (when real-world data changes) and safeguards for misuse or failures.

Data Flow Consider a supervised image classification example (cats vs. dogs): images and labels enter the pipeline; images are preprocessed; the model predicts labels; a loss function compares predictions to true labels; the optimizer updates the model to reduce the loss; the process repeats over many examples until performance stabilizes; finally, the trained model predicts on new images. In unsupervised customer segmentation, customer records flow into a clustering algorithm; the algorithm groups similar customers; analysts inspect and label clusters by behavior; marketing uses these groups for tailored campaigns. In RL, an agent observes its state (like a game board), selects an action, receives a reward and new state, updates its policy, and repeats, learning strategies that increase total rewards.

Supervised Learning: Technical View

• Inputs and Labels: Each training example pairs features (inputs) with a correct label (output). Features for images are pixel values; for text, token IDs; for tables, numeric or encoded categorical fields.
• Model: A function that maps inputs to outputs. It has parameters that adjust during training to fit the data.
• Loss Function: A formula that measures how wrong the model is. For classification, cross-entropy loss is common; for numeric prediction, mean squared error is common.
• Optimization: An algorithm (often gradient descent and its variants) that tweaks parameters to reduce the loss by following the slope of the error landscape.
• Generalization: The real goal is to perform well on new, unseen data. Techniques like validation sets, regularization, and early stopping help avoid overfitting (memorizing training data).

Unsupervised Learning: Technical View

• No Labels: The algorithm must infer structure without answer keys. Clustering (like k-means) groups points by similarity; dimensionality reduction (like PCA) finds key directions that explain variance.
• Similarity: A distance measure (like Euclidean or cosine) defines how close points are. The choice of distance affects cluster shapes and results.
• Use Cases: Market segmentation, anomaly detection, document grouping, and feature learning. Insights often guide business decisions or prepare data for supervised models.

Reinforcement Learning: Technical View

• Agent-Environment Loop: The agent sees a state, takes an action, receives a reward and the next state, and repeats. The goal is to learn a policy that maximizes expected long-term reward.
• Exploration vs. Exploitation: The agent must try new actions (exploration) to learn, but also use known good actions (exploitation) to get rewards now. Balancing the two is central to RL training.
• Credit Assignment: Rewards may come much later than the actions that caused them. Algorithms like temporal difference learning help assign credit to earlier steps.
• Applications: Games (chess, Go), robotics control, resource allocation, and dynamic recommendations.

Natural Language Processing: Technical View

• Representation: Text is converted to numbers so computers can process it. Tokenization breaks text into pieces (words or subwords), which are mapped to IDs.
• Understanding and Generation: Models learn to predict missing words, classify sentiment, answer questions, or translate between languages. They learn patterns in word order and meaning.
• Data and Ambiguity: Human language is messy and context-dependent. Good datasets and careful evaluation are needed to avoid misunderstanding and bias.
• Applications: Chatbots, virtual assistants, translation tools, document summarizers, and content moderation.

Computer Vision: Technical View

• Representation: Images are grids of pixel values (e.g., red, green, blue intensities). Preprocessing may include resizing, normalization, and augmentation (like random crops) to improve robustness.
• Tasks: Classification (what is in the image), detection (where objects are), segmentation (which pixels belong to what), and tracking (how things move over time).
• Data Demands: Vision models often need many images to learn well. Careful labeling and diverse samples reduce bias and improve performance in the real world.
• Applications: Self-driving perception, medical image analysis, security, manufacturing quality control.

Choosing the Right Approach

• If labels are available and the goal is prediction, use supervised learning. If the goal is to explore structure or group items without labels, use unsupervised learning. If actions lead to future rewards and sequences matter, use RL.
• For language tasks, use NLP methods that understand order and meaning in text. For images and video, use computer vision methods that exploit spatial structure.
• Always consider data availability, quality, and cost of labeling when selecting an approach.

Evaluation and Metrics

• Accuracy measures how often predictions are correct, but it is not always enough. Precision and recall matter when class balance is skewed (like rare fraud cases). For ranking or recommendation, metrics like top-k accuracy or mean average precision are used.
• Beyond numbers, evaluate fairness: do errors harm certain groups more? Also test robustness: does performance hold up under small changes or noisy inputs? Safety testing checks for harmful failure modes.

Risks and Safeguards in Practice

• Job Displacement: Plan for human-AI collaboration; automate tasks, not people. Provide training and transition support.
• Bias: Use diverse, representative datasets; audit models for disparate impact; document data sources and model limits.
• Malicious Use: Implement access controls, logging, and monitoring. Follow laws, ethical guidelines, and internal review processes.
• Governance: Establish clear roles for oversight, red-teaming (testing against misuse), and incident response. Regularly update models as data and conditions change.

Data Lifecycle and Monitoring

• Data Drift: Over time, the world changes (new slang, new fraud patterns). Monitor model inputs and outputs to catch drift early.
• Feedback Loops: User behavior can shift because of model decisions (e.g., showing certain recommendations changes future clicks). Design experiments and checks to prevent harmful loops.
• Retraining: Schedule periodic retraining with fresh data, and validate that the new model truly improves outcomes.

Deployment Considerations

• Latency: Real-time applications need fast predictions; batch tasks can be slower. Match model size and infrastructure to speed needs.
• Reliability: Add fallbacks and human-in-the-loop review for high-stakes decisions. Log decisions and reasons when possible for accountability.
• Security and Privacy: Protect data at rest and in transit; anonymize or minimize data collection; comply with regulations.

Putting It All Together A practical AI project starts with a problem (e.g., detect fraud), collects and prepares data (transaction histories), selects supervised learning, trains and evaluates a model (optimizing for recall while keeping false positives manageable), and deploys it into the payment system with monitoring. For an unsupervised project (customer segmentation), data is clustered to find natural groups, then marketing designs personalized campaigns. For an RL project (delivery route optimization), an agent simulates routes, gets rewards for faster on-time deliveries, and learns strategies that handle traffic patterns. Across all projects, ethical review, fairness testing, and safeguards remain central to responsible AI.

This lecture builds a clear, complete picture of AI: what it is, how it is built, where it is used, and why ethics matter. It starts with a broad definition—machines doing tasks that would need intelligence if a human did them—and distinguishes today’s narrow AI from the still-hypothetical general AI. It explains machine learning as the main engine of modern AI and breaks it into supervised learning (labels teach predictions), unsupervised learning (patterns emerge from unlabeled data), and reinforcement learning (agents learn by rewards). It introduces two core application areas—natural language processing for understanding and generating human language, and computer vision for interpreting images and video—then connects these methods to real-world uses across healthcare, finance, transportation, entertainment, retail, and manufacturing. The lecture closes with key risks: job displacement, bias from flawed data, and malicious uses like autonomous weapons or surveillance, urging ethical safeguards and responsible deployment.

To practice, start by identifying a small, real problem you care about—such as filtering spam, grouping similar customers, or predicting demand—and choose the matching approach (supervised, unsupervised, or RL). Gather clean, representative data; split into training and testing sets; pick simple, well-documented models; and measure results with the right metrics. Add checks for fairness and robustness, and outline how you would monitor a model in the real world. Try building a toy NLP task (like intent classification) or a simple vision classifier (like cats vs. dogs) to make the ideas concrete.

For next steps, learn more about data collection and labeling, evaluation metrics, and model monitoring. Explore domain-specific topics: tokenization and text pipelines for NLP, image preprocessing for vision, and policy learning for RL. Read about AI ethics, bias auditing, and governance frameworks to build systems that are not only accurate but also fair and safe. Expanding your toolkit while deepening your ethical awareness prepares you for real projects.

The core message is balance: AI is powerful and already changing many parts of life, but it must be guided by care, oversight, and respect for people. Choose the right method for the problem, build with quality data, test fairly, and design safeguards from the start. With these habits, you can use AI to create helpful, trustworthy systems that improve lives while minimizing harm.