By scaling deep learning, the AI industry has figured out recipes for training outstandingly general and capable foundation models. As these models improve, they become not only excellent general-purpose assistants but also invite building new software systems that are defined in part using inherently fuzzy natural-language specifications. For the development of such AI systems, the bottlenecks are increasingly less centered on raw capability for ad-hoc requests but instead on systems engineering concerns like specification, reliability, efficiency, and transparency.
Omar Khattab,
MIT
In this talk, I will describe our work on models, abstractions, algorithms, and application systems that seek to make the design and implementation of AI systems more of a scientific discipline. This comes together under the two themes of Declarative AI Programming and Self-Improving AI Systems, unified into the DSPy paradigm for AI programming and supported by algorithms for inference and learning for foundation models, most recently Genetic-Pareto prompt optimization (GEPA), multi-module GRPO (mmGRPO), and Recursive Language Models (RLMs).
Omar Khattab is an assistant professor at MIT EECS and a member of CSAIL. His research creates models, algorithms, and abstractions for building reliable and scalable AI systems. He authored the ColBERT retrieval model, which has helped shape the modern landscape of neural information retrieval, and the DSPy framework for building and optimizing LLM-based systems, both of which are downloaded millions of times per month and have sparked applications at dozens of organizations. His Ph.D. research was supported by the Apple Scholars in AI/ML Ph.D. Fellowship and his work has received a SIGIR Best Paper Award.