John Langford, the machine learning researcher who led AI Frontiers inside Microsoft Research, says the group no longer exists and that he has left. He laid it out on 6 November in a post on his blog, Machine Learning (Theory), at hunch.net. The team, about 50 people by his count, was folded into Microsoft AI in early September, and his account of what came next reads like a resignation letter. It carries no response from Microsoft.

He says the group formed in 2023, after ChatGPT arrived, because a lone researcher’s idea tended to land late and look weak next to what better-resourced labs were running. Several colleagues felt the same, so they built a sub-organization to run research as coordinated efforts. Compute was often a severe bottleneck, he writes, and disagreements were sometimes hard to settle.

On the output side, he points to agent work: Autogen, a framework for orchestrating agents, and Fara, an agent that operates a computer. A smaller set of researchers in New York worked on basic capabilities: a training objective called Next Latent, an improved version of the Muon optimizer called Dion3, and the Full Bandwidth Transformer, which uses recursion to get more out of a model. The group also kept a shared code repository open to all of Microsoft Research, including a pretraining codebase for small models that he described in a COLM keynote last month. He says he knows of at least four separate attempts inside the lab to build a model-making pipeline.

Then came the move into Microsoft AI, which is much larger. Langford grants that merging two AI organizations can make sense on paper. In practice, he says, the leadership there, who came from Inflection, believes in scaling and “aggressively not research.” He concedes the business logic: returns to scale are far easier to predict than returns to research. But the work in that setting is mostly about getting and refining data, a long way from what his people did, and he says there was plenty of evidence the new organization did not want them.

His counterargument rests on one result. Langford says experiments at ByteDance hint that a version of the Full Bandwidth Transformer could rival a model twice its size. If that holds, the saving on training and inference would be large, and he asks what halving those costs is worth when the industry has put close to a trillion dollars into the field. The post does not link the experiments or give their numbers, and the claim comes from someone with a stake in the architecture, so it stands as a hypothesis.

Langford says he found homes for his team as best he could and left. He writes that every other founding member of the AI Frontiers leadership team appears to have reached the same conclusion.

He also explains why he did not return to Microsoft Research. One reason is that he does not want to work where research has no route into products, and his attempts to get ideas considered inside Microsoft AI went nowhere. The other is what he calls a broken bargain, the promise that strong work gets rewarded. In his telling, the group did good work and was reorganized away without its consent. He expects more of the same “until MSR leadership develops a will to effectively organize to meet the moment.” It is a pointed public charge against his former employer’s research leadership, and it is his opinion alone.

The post also says there will be no AI Frontiers internships next year. Autogen and Fara both came out of the group, and Langford does not say who maintains them now, which is the first thing to ask if your product depends on either.

Reported by Machine Learning (Theory) (hunch.net), John Langford’s blog, on 6 November 2026.