Overview
For the first time in computing, we can specify arbitrary intelligent behavior directly in natural language, thanks to language models. But can we use this to build systems that are not only intelligent but also reliable, efficient, and maintainable? This course's answer is yes, and it seeks a principled approach to engineering such systems.
We start by learning why this is difficult. Language models are grown from large piles of broad data rather than engineered around any formal specification. In contrast, every system we build has specific stakeholders, whose situated needs create elaborate requirements we have to meet. As AI engineers, we thus have to elicit these requirements and express them as declarative specifications that endure as models change. Moreover, we have to design systems that make the most of model capabilities under a budget, that learn from sparse and qualitative stakeholder feedback, and whose essential properties are formally guaranteed or empirically demonstrated.
Throughout the course, we consider trade-offs between reliability, efficiency, and maintainability and seek out durable principles that outlast individual generations of models. Students construct practical, open-ended, and competitive systems in the homeworks and project.
Course information
- Meetings
- Wednesdays, 11:00am–1:00pm
- Room
- 32-144
- Instructor
- Omar Khattab
- TAs
- Jyo Pari and Zekai Wang
- Office hours
- Zekai: Mondays, 1–2pm, room 45-746
Jyo: Thursdays, 1–2pm, room 45-746 - Discussion
- Piazza
- Course materials
- Canvas Modules
- Units
- 2-0-10, graduate
- Satisfies
- AUS, AAGS in Computer Systems and Artificial Intelligence
Prerequisites
Prerequisites are 6.3900 and either 6.1020, 6.1800, or 6.1810, or permission of instructor. We recommend 6.1210 or equivalent algorithmic maturity.
Course structure
We meet once a week for two hours. Students will complete two individual homeworks and a substantial project. Five short quizzes will be given in class, with the best three counting toward the grade.
Models and compute
We thank Prime Intellect, Thinking Machines, and Mixedbread for their generous compute support. Thanks to them, the course provides the model endpoints and GPU compute that the homeworks and the project require.
