Incoming Assistant Professor of Electrical Engineering and Computer Science, MIT
Talking with Computers to Solve Bigger Problems
In manufacturing, innovation leads to people spending less time doing manual labor and repetitive work. The work that Omar Khattab (CMU 2019) does with large language models (LLMs) is doing something similar for computing. His research looks at how we talk to computers and what they do with what we tell them. If he can improve our conversations with them, computers could take on more work and jumpstart human problem-solving.
“Improving LLMs and making them a reliable component in larger software systems would allow people to solve bigger and harder problems,” Omar says. “When people have a good idea in their domain, they should find it much easier to build smart software that solves their problems more directly.”
The conversation around processing natural language was already changing at the time Omar began his undergraduate studies in computer science at CMU Qatar. He just completed his Ph.D. and will continue studying LLMs when he joins the electrical engineering and computer science department at MIT as an assistant professor this fall.
The impact of innovative LLMs is becoming more evident with the appearance of AI overviews in search engine results. Searches used to just match to keywords, but now they’re using some of the frameworks that Omar works on to find much better results, summarize search topics, and deliver a more sophisticated answer to searched terms and questions.
Omar also studies how LLMs can become a part of the way we program software applications. Conventionally, programming languages and their tools like compilers require programmers to specify every little detail. This can be hard when trying to build higher-level software that interacts with users in natural language.
“Right now, people mostly write software using low-level building blocks. My work considers how we can raise this level of abstraction and allow experts in domains like education or healthcare to describe the behavior of their software systems using higher-level constructs, partly specified in a human language,” he says.
Story by Elizabeth Speed