A Conversation with Mozilla’s Chelsea Troy

Do large language models make developers 26% more productive, as a paper recently claimed? Chelsea Troy, leader of Mozilla’s MLOps team and a lecturer in computer science at the University of Chicago, isn’t so sure. There’s so much hype right now—we need voices like Chelsea’s to cut through the noise and help us think a little bit more deeply about what programming actually requires. And that’s just what she did on this episode of Live with Tim O’Reilly.Chelsea joined Tim for a conversation on why widely cited productivity studies like the “26% paper” are fundamentally flawed, how the real work of programming—reading and analyzing existing code rather than writing from scratch—hasn’t changed, and what skills developers actually need to thrive in an AI-augmented world. Above all, they thought through the challenge of how developers can harness AI’s power without becoming overly reliant on it, which could erode the very cognitive skills that make them effective problem-solvers in the first place.It’s no surprise that human creativity remains irreplaceable, as Chelsea made clear: Even when assisted by AI, programmers must rely on their investigative abilities to ask the right questions, on their evaluative skills to assess solutions, and on their innovative thinking to push beyond existing patterns. “Machine learning models, the way they are built, the way they are architected, the way they are trained, the way they fundamentally work is to replicate patterns,” she explained. “They’re not designed to generate new patterns, which is exactly what you need in order to move the state of the art.”