John Langford, one of the founders of Microsoft Research's AI Frontiers group, says the group is finished and that he has left Microsoft. In a post on his long-running Hunch.net blog dated November 6, he writes that AI Frontiers was reorganized into Microsoft AI (MAI) in early September, that its researchers were moved without their consent, and that there will be no AI Frontiers internships next year.
Why it matters: AI Frontiers produced some of Microsoft's best-known open AI work, including the AutoGen agent framework. Langford's account is an unusually direct, first-person description of a tension running through every big AI company right now: whether money and compute go to scaling up today's models or to research on new ideas.
What AI Frontiers was
Langford says he and others set up AI Frontiers inside Microsoft Research after ChatGPT's release, concluding that the lab needed more organized efforts to keep up with well-resourced rivals. The group sat between top-down and bottom-up research direction. By his account:
| Size | About 50 people |
| Biggest constraint | Compute, which he says was often a severe bottleneck |
| Agent work | AutoGen (agent orchestration), Fara (computer-use agent) |
| Fundamental research (New York) | Next Latent objective, Dion3 (an improved Muon optimizer), the Full Bandwidth Transformer |
| Infrastructure | A monorepo shared across Microsoft Research, including an edge-scale pretraining codebase featured in his COLM keynote |
What happened, in Langford's telling
In early September the group was moved into Microsoft AI, the larger organization run by leadership that came over from Inflection. Langford says MAI's leaders prioritize scaling over research. StartupHub.ai, summarizing the post, notes he called that scaling-only view misdirected.
He gives three reasons for leaving:
- He could not find a way to get research into real use at MAI.
- He doubts Microsoft Research can currently offer that path either.
- He worries that Microsoft Research's basic promise, roughly that good work leads to good outcomes, has broken down.
He says every other founding AI Frontiers leader reached a similar conclusion, and that he placed his team members as best he could before leaving. He does not name a new employer.
Langford argues research still pays off, pointing to ByteDance experiments suggesting a variant of the group's Full Bandwidth Transformer can match a model twice its size. StartupHub.ai adds that he knew of at least four separate model-building efforts inside Microsoft, only one of them shared.
We found no public comment from Microsoft, and StartupHub.ai's report includes no statement from the company.
Who should care
- AutoGen and Fara users: nothing in the post says these projects are being shut down, but the team that built them is being dispersed. Watch their repositories for maintainer changes and release cadence before betting new work on them.
- AI researchers and students: the loss of AI Frontiers internships removes one industry route for research-focused placements next year.
- Anyone tracking Microsoft's AI strategy: the post fits a pattern of Microsoft concentrating AI work under Microsoft AI. Langford's closing line is that Microsoft could become attractive for research again if its leaders develop more wisdom.
What to watch
Where Langford and the other AI Frontiers founders land, whether Microsoft responds publicly, and whether the Full Bandwidth Transformer work continues elsewhere.