Microsoft Research's AI Frontiers lab has joined Microsoft AI, and machine-learning researcher John Langford says the shift exposed a dispute over scaling models versus pursuing research.
In a post on his blog, Langford said he helped organize AI Frontiers in 2023 to give researchers a way to tackle work that could be difficult for individuals competing with larger, better-supported groups. He estimated the team had about 50 people. Langford wrote that he had “situated my folks” and left Microsoft after AI Frontiers moved into Microsoft AI, rather than returning to Microsoft Research.
He said combining the groups made sense in theory but that Microsoft AI's former Inflection leadership had a different approach. “The core philosophy there is scaling, and aggressively not research,” Langford wrote. He acknowledged that larger models can improve predictions. From a business perspective, he said, returns from scaling are more predictable than returns from research.
Microsoft Research's AI Frontiers page, updated Sept. 14, says the lab joined Microsoft AI and would continue agent and model research there. The page lists Ece Kamar as corporate vice president and lab director, and names AutoGen and Fara among the lab's projects. Langford described AI Frontiers as “over,” but his post does not specify whether he meant the original group he helped organize or the broader lab Microsoft says will continue.
Langford also co-authored the Full-bandwidth transformer paper. It reports that one-billion-parameter models trained on up to 400 billion tokens performed as well as or nearly as well as standard transformers trained on about 50% more tokens.
Microsoft has described Langford as one of the founding members of Microsoft Research New York and the principal developer of Vowpal Wabbit, whose first public version appeared in 2007. The International Conference on Machine Learning's 2021 board lists him as president.
