Omar Khattab

He graduated from Carnegie Mellon UniversityCarnegie Mellon UniversityCarnegie Mellon University (CMU) is a private research university in Pittsburgh, Pennsylvania. It grew out of the Carnegie Technical Schools, founded with funding from Andrew Carnegie in 1900, which were renamed the Carnegie Institute of Technology in 1912 and took the present name after merging with the Mellon Institute in 1967. The university is known for computer science, robotics, engineering, and the arts, and has campuses in Qatar, Silicon Valley, and Rwanda. As of September 2025, it enrolled 16,582 students. Its current president is Farnam Jahanian. On September 30, 2026, the university announced a $3 billion gift from Ken Griffin and plans to build a new campus in Miami.Open the full entry’s Qatar campus in 2019, where he studied computer science.⁠[1]

In 2020, he and Matei Zaharia published ColBERT, which encodes queries and documents separately before fine-grained late interaction. Precomputing document representations makes BERT-based semantic matching more efficient for retrieval.⁠[2]

In an article with Christopher Potts and Matei Zaharia for Stanford HAI, he advocated neural information retrieval as a foundation for better search, retaining the documents behind answers and incorporating new knowledge through index updates.⁠[3]

From June to September 2022, he was a PhD research intern at AppleAppleApple is a public technology company founded in 1976 and headquartered in Cupertino, California, United States. Its business includes the iPhone, Mac, iPad, wearables, and services such as digital content, payments, and cloud storage. The company builds its product system through in-house hardware, operating systems, and services.Open the full entry.⁠[4]

The DSPy paper, on which he is first author, presents language-model applications as composable declarative modules. Compilation and optimization improve model-call pipelines against task metrics, reducing reliance on manually maintained prompts.⁠[5]

A Two Sigma article he co-authored organizes language-model software into seven layers and compares five framework categories by reliability, performance, portability and extensibility, emphasizing separation of responsibilities between layers.⁠[6]

He co-authored The Shift from Models to Compound AI Systems. The article examines combinations of models, retrieval and tools, treating quality, cost and optimization as system-wide concerns.⁠[7]

From June 2024 to June 2025, he was a research scientist at Databricks, working on natural language processing, information retrieval and machine-learning systems.⁠[4]

In a personal essay, he recommends organizing research around sustained projects rather than individual papers, choosing problems with room to develop, and building impact through rapid iteration, usable software releases and community feedback.⁠[8]

He began a computer-science PhD at StanfordStanford UniversityStanford University is a research university in California founded by Leland and Jane Stanford in 1885 and opened in 1891.Open the full entry in 2019 and completed it in January 2025, advised by Christopher Potts and Matei Zaharia. His doctoral work focused on foundation-model programming for more reliable and scalable AI systems.⁠[4]

He co-authored WARP, a multi-vector retrieval engine, with Jan Luca Scheerer, Matei Zaharia, Christopher Potts and Gustavo Alonso. The paper received the SIGIR 2025 Best Paper Award.⁠[9][10]

Carnegie Mellon named him to its 2025 Tartans on the Rise alumni list. Its profile highlighted his work on higher-level language-model programming abstractions that help domain experts build software.⁠[1]

In an AI Engineer talk, he argued for separating application specifications, control flow, tools and evaluations from changing models and optimization methods. He emphasized defining the problem a system must solve and avoiding designs tied to temporary prompting tricks.⁠[11]

He began as an assistant professor at MIT in July 2025. His current appointment is in Electrical Engineering and Computer Science, where he also holds the TIBCO Founders Career Development Professorship.⁠[4][12]

He contributed to GEPA, which uses natural-language reflection on execution traces and feedback to generate and select prompt variants. Its comparisons concern particular tasks and do not establish that prompt optimization universally outperforms reinforcement learning.⁠[13]

He and Alex Zhang introduced recursive language models, storing long inputs in an executable environment where a model can inspect and partition them with code and call sub-models on local questions. Early experiments used limited recursion depth and focused on long-context tasks.⁠[14]

Pedagogical RL, which he co-authored, uses information available to a teacher but hidden from a student to generate teaching trajectories while considering whether the student can learn them. The article reports preliminary task experiments on learning from better-targeted demonstrations.⁠[15]

He, Jacob Xiaochen Li and Rick Battle formulated machine studying: agents study a corpus without knowing future questions, then solve tasks more efficiently while retaining access to the material. The work measures expertise through the relationship between quality and inference cost and introduces the developing StudyBench benchmark.⁠[16]

He and Alex Zhang studied how agent harnesses can help models transfer from shorter to longer tasks, focusing on context offloading and programmatic sub-calls. Their recursive-language-model experiments examine the relationship between task decomposition and the training distribution.⁠[17]

An MIT profile describes his OASYS lab’s aim as addressing AI’s last mile: adapting general models to specific users’ needs. Khattab emphasizes reusable open-source systems methods rather than solutions confined to a single vertical application.⁠[18]

He teaches MIT’s 6.S978, Engineering AI Systems and Agents, covering reliability, efficiency and maintainability through declarative specifications, evaluation, feedback and compute budgets.⁠[19]

In a personal X post, he said AI productivity gains had not made his research work meaningless. Better tools expanded his ambitions, so high-quality results still required time and expertise. He also expressed concern about concentrated power. These statements describe his own research experience and views.⁠[20]

He, Rikiya Takehi, Jayden Teoh and Alex Zhang published experiments in which language models directly read and modify other neural networks’ weights. The article reports early successes alongside experiments without meaningful gains; the results do not establish a general replacement for model training.⁠[21]

Personal website: Official website: https://omarkhattab.com/ (opens in a new window). X: Public account: https://x.com/lateinteraction (opens in a new window). GitHub: Public account: https://github.com/okhat (opens in a new window).⁠[22]