Schedule
| Date | Topic |
|---|---|
| From foundation models to AI systems | |
| Sep 9Lecture 1 | Language Models as a Computational PrimitiveSlides (PDF) · Recording For the first time in computing, it's possible to embed arbitrary intelligent functions inside software programs, by providing natural language specifications to language models. We will study how language models are "grown" rather than engineered and why it thereby remains difficult to build reliable, efficient, and maintainable software systems around them. |
| Sep 16Lecture 2 | Requirements for Situated AI SystemsSlides (PDF) · Recording · HW1 public overview (PDF) Language models supply broad capabilities, but each system we build will have specific stakeholders with elaborate requirements. Thus, as AI engineers, we must elicit these requirements and express each in the shape that fits it: natural language for fuzzy rules, code for symbolic restrictions, and examples for the long tail of exceptions. Across several design tensions, from decomposition to security, we will then weigh how much to let the model do against how much to engineer around it. |
| AI programming and agents | |
| Sep 23Lecture 3 | Declarative AI ProgrammingSlides (PDF) · Recording · Quiz 1 (PDF) The language model we use changes every month. How can we write a durable AI program that keeps meeting its requirements despite that? We will ask when hand-tuning for today's model is premature optimization, and how a higher-level programming model avoids it. AI programs become code that composes typed natural-language signatures, which decouple what we want from how any particular model should do it. |
| Sep 30Lecture 4 | Agent DesignSlides (PDF) · Recording We will build on the previous lecture by defining the semantics of a declarative AI program. We will then study agents: AI programs that loop and assign the model consistent goals and the flexible control across many calls required to pursue them. We then ask how the program should process the current state of the environment for the model to observe, and how it should construct the action interface of the model. |
| Evaluation and evidence | |
| Oct 7Lecture 5 | Learning from HW1 SystemsRecording · Quiz 2 (PDF) |
| Oct 14Lecture 6 | Retrieval and Reasoning over Massive ContextsSlides (PDF) · Recording Some tasks require our models to deeply reason about evidence spread across an entire collection of documents. We will consider the difficult trade-offs this poses and discuss the retrieval and long-context processing paradigms developed to tackle it. We will end with retrieval problems that remain difficult even in 2026. |
| Oct 21Lecture 7 | Evaluating AI SystemsSlides (PDF) · Recording · Quiz 3 (PDF) Having built many systems and agents by this point, how do we know whether they work? We will start from qualitative coding and then discuss when a formal eval is required. We will also consider the range of useful evals, some of which consider average behavior like ML benchmarks whereas others represent something closer to software assertions. We will cover also the minimal indispensable statistics for useful evals. |
| Learning from evidence | |
| Oct 28 | Project and Course ConsultationsIndividual and team consultations on projects, lectures, quizzes, and homework. |
| Nov 4Lecture 8 | Downstream Learning for AI SystemsSlides (PDF) Engineers often improve an AI system by hand-crafting better prompts and programs. We will study learning algorithms that search over instructions, demonstrations, and the program structure itself from downstream rewards and even from qualitative feedback. We will consider how such algorithms respect the declarative programming model from earlier lectures. |
| Nov 11 | No classVeterans Day. |
| Nov 18Lecture 9 | Downstream Learning in WeightsThere are cases where a situated AI system strikes its best trade-offs only by updating the model weights from downstream feedback. We will compare methods for supervised fine-tuning, reinforcement learning, and distillation and develop intuition for cases where the weights are the right parameters to learn vs. the program around the model. |
| Later topics and projects | |
| Nov 25 | Project Presentations I |
| Dec 2 | Project Presentations II |
| Dec 9 | Project Presentations III |