Microsoft Research's AI Frontiers group is over, says a founding leader

In a first-person post on hunch.net, one of the founding leadership team of AI Frontiers (AIF), a sub-organization of Microsoft Research (MSR), explains why the group no longer exists. The post does not name its author.

The author says that after ChatGPT's release in 2023 it was clear to him that research had to be more organized to succeed than was typical inside MSR. Other research groups were applying so much support and focus that one person with a clever idea was likely to arrive late and unconvincing: one person could not move fast enough, and the baselines one person could assemble were not strong enough to show an idea was worth pursuing. Several colleagues thought the same, and together they created AI Frontiers.

AIF tried to organize coherent research efforts, in what the author calls the messy middle between top-down and bottom-up. There were disagreements, and compute was often a severe bottleneck. Even so, he writes, he had the general impression that the group was doing pioneering things, especially given its small size (about 50 people). Much of its work went into making agents work, for example the Autogen orchestration framework and the Fara computer use agent. In New York, some researchers worked on new fundamental capabilities: the Next Latent objective, which enables extrapolation; Dion3, an optimized or improved Muon; and the Full Bandwidth Transformer, which makes models function substantially more powerfully via recursion.

AIF also shared its work with the rest of MSR. It built a monorepo shared by default with all of Microsoft Research. That repo held the proven-capable edge-scale pretraining codebase the author discussed in a COLM keynote the previous month. He says creating a well-formed model-building pipeline takes substantial effort, and that he knows of at least four efforts in MSR trying to build such pipelines; the other three are not in repos visible within MSR.

In early September, AIF was reorganized into Microsoft AI (MAI). The author concedes this could make sense on the surface: combining two AI organizations allows better concentration of effort, AIF was ahead of MAI in some places and MAI ahead in others, and MAI is much larger. In practice he found the philosophy of MAI's ex-Inflection leadership incompatible with AIF's: "scaling, and aggressively not research." He grants there is plenty of evidence that bigger models are better models, and that from a business perspective returns to scale are far more predictable than returns to research. In a scaling-only environment, he says, the main effort goes into getting, making and refining data, which is a long way from what AIF did. Beyond the mismatch in focus, he writes, there was plenty of evidence that MAI did not really want them.

He calls the scaling-only focus both misdirected and not fun, though an understandable failure mode, since scaling is straightforward and people often resolve a tradeoff between ideals like innovation and size by adopting one and dropping the other. Returns to research, he argues, are real and significant. His example: Bytedance experiments suggest a variant of the Full Bandwidth Transformer might be comparable to a model twice the size, which would mean significant savings in training and inference for the same capability. He asks readers to work out what halving training and inference cost is worth when the industry has investments approaching a trillion dollars.

He settled his team as best he could and left, and says he is fortunate that many other people can see the value of research and innovation. Every other founding AIF leadership team member seems to have reached a similar conclusion.

Why not return to MSR? He gives two obstacles. First, he is not comfortable working where his research has no path to real use; he could not get innovation into use, or even seriously considered, while in MAI, so MSR's prospects look small. Second, he fears a breakdown of the basic deal in MSR, "do good work and good things happen": as far as he can tell, AIF did good work and was non-consensually reorganized away. He would not be surprised if such reorganizations continue until MSR leadership develops a will to organize effectively for the moment.

His verdict: AI Frontiers is over. There will be no internships with its people in the coming year. He closes with best wishes for those he leaves behind, and says it is not hard to imagine a Microsoft that becomes attractive for research again if the people in control develop a little more wisdom.

Key facts

• A founding leader of AI Frontiers, a Microsoft Research sub-organization of about 50 people, says it was reorganized into Microsoft AI in early September and is now over.
• He says MAI's ex-Inflection leadership holds a philosophy of "scaling, and aggressively not research", incompatible with AIF's work, and that there was plenty of evidence MAI did not really want AIF.
• AIF's output included the Autogen orchestration framework, the Fara computer use agent, the Next Latent objective, Dion3 (an improved Muon) and the Full Bandwidth Transformer.
• He argues research pays: Bytedance experiments suggest a Full Bandwidth Transformer variant might match a model twice the size, against industry investments he puts near a trillion dollars.
• He left and says every other founding AIF leader seems to have done the same; he does not want to return to MSR because research there has no path to real use and the basic deal has broken down.

Why it matters

This is a first-hand account of how a small research group inside a large AI company was absorbed by a bigger product-focused organization. The author frames it as a clash between two bets: scale up models and data, or keep investing in research that might make models cheaper to build and run. His concrete case for research is the Full Bandwidth Transformer, where Bytedance experiments suggest a variant might match a model twice the size, which would cut training and inference costs for the same capability. The post also records the founding rationale of AIF: after ChatGPT in 2023, a lone researcher could no longer build strong enough baselines or move fast enough.

Who it affects

Most directly, the roughly 50 people of AI Frontiers, including the founding leadership team, who the author says seem all to have left, and the interns who will not be hosted by the group in the coming year. It also touches Microsoft Research staff, who now face the author's concerns about a breakdown of the deal that good work leads to good things, and anyone who relied on AIF's work such as Autogen or Fara. The source does not say what happens to those projects.

How to use it

There is nothing to install or adopt. Read it as a case study: the author's account of why AIF organized in 2023, how it shared a monorepo with all of MSR, and what went wrong when a research group was merged into a much larger scaling-focused organization. Researchers weighing similar structures can take the author's stated reasons for the founding, and his two reasons for not going back to MSR, as one practitioner's view.

How solid is it

This is a single personal account from someone directly involved, published on hunch.net, and it is the author's interpretation throughout. The text does not name him, give a date for the post, or carry any statement from Microsoft, MSR leadership or Microsoft AI. The claims about MAI's philosophy, about MAI not wanting AIF, and that other founding leaders reached the same conclusion ("seems to") are his own. The Bytedance result is given only as "suggest" and "might", with no metrics or details.

Risks and caveats

The post is one side of a reorganization, written by someone who left, and it does not say how many AIF people were laid off, moved or departed. His forecast that such reorganizations may keep happening in MSR is a personal expectation, not a fact. The twice-the-size claim for the Full Bandwidth Transformer is hedged and unsupported by numbers in the post. He also concedes that bigger models are better and that returns to scale are more predictable for business than returns to research, so his argument rests on the case that research savings are large enough to matter.

“The core philosophy there is scaling, and aggressively not research.”

— Founding AI Frontiers leader, on Microsoft AI, hunch.net