A founding leader says Microsoft's AI Frontiers lab is over

John Langford, a machine-learning researcher who writes the blog Machine Learning (Theory), posted that “AI Frontiers is over”. He described AI Frontiers as a sub-organisation of Microsoft Research that he and others set up in 2023, after the release of ChatGPT, to run more coordinated research than individual projects allowed. He put its size at about 50 people. By his account it worked on agents, including the AutoGen orchestration framework and the Fara computer-use agent, and on newer ideas such as new training objectives and architectures.

The reorganisation itself is documented by Microsoft. A note on the lab’s own page, dated 14 September 2026, says Microsoft Research AI Frontiers is now part of Microsoft AI, and that the group will pursue the same goal of frontier agentic and model research there, adding that more would be said about the transition. Langford writes that the group was moved into Microsoft AI in early September.

What he disputes is the outcome. He says the philosophy of the Microsoft AI leadership, which he describes as ex-Inflection, is “scaling, and aggressively not research”, that he could not get the group’s work considered for use there, and that “there was plenty of evidence that they just did not really want us”. He argues that returns to research are real and gives, as an example, ByteDance experiments suggesting that a variant of one architecture from his group might match a model twice the size. He also says he considered returning to Microsoft Research but was put off by two obstacles: no path from research to real use, and what he sees as a breakdown of the basic deal there. Langford says he left, that every other founding member of the leadership team appears to have reached a similar conclusion, and that there will be no internships with the group next year.

This is a first-person account by someone who left, and it is the only source for the departures and the reasons. No Microsoft statement responding to it, and no independent reporting, appeared in the sources checked. The characterisation of Microsoft AI’s philosophy is Langford’s. Microsoft AI has meanwhile been shipping its own models, including MAI Thinking 1 and, in October, MAI Base 2, which is consistent with the emphasis he describes but does not confirm it.

It is a data point in the wider competition for AI research talent and in the argument over whether progress now comes mainly from scale or from new methods.