John Langford says Microsoft's AI Frontiers is over, and blames a scaling-only Microsoft AI

On November 6, John Langford, one of the best-known researchers at Microsoft Research, wrote on his blog that AI Frontiers, the roughly 50-person group he helped create there in 2023, has ended. The group was moved into Microsoft AI in early September. Langford says he placed his team as best he could and left, and that every other member of the founding leadership “seems to have come to a similar conclusion.”

This is one participant’s account. We found no Microsoft statement confirming or disputing it, and Langford names no other departing staff. It still deserves attention, because it comes from a senior insider and makes a specific, testable argument about how a company with enormous AI spending is choosing to allocate research effort.

What AI Frontiers was

Langford’s post explains why the group existed. After ChatGPT, he concluded that a clever idea pursued by one researcher in the traditional Microsoft Research style would tend to arrive late and with baselines too weak to convince anyone. AI Frontiers was formed as a sub-organization inside MSR to run coordinated efforts, somewhere between top-down and bottom-up. He names compute shortages as an often severe bottleneck.

Most of its work went into making agents function, including the AutoGen orchestration framework and the Fara computer use agent. A group in New York pursued more basic ideas: the Next Latent objective, the optimizer Dion3, and the Full-Bandwidth Transformer. AI Frontiers also built a monorepo shared by default across MSR, which held a pretraining codebase Langford discussed in a COLM keynote last month. He treats that as a real asset and says he knows of at least four separate model-building pipelines at the company, three of them not in repositories open within MSR. That duplication is a small but telling detail about how large labs actually operate.

The argument: scale versus research

Langford concedes that the merger made sense on paper, since each organization was ahead of the other in places and Microsoft AI is much larger. His objection is to its leadership, which he describes as ex-Inflection. Microsoft hired Inflection co-founders Mustafa Suleyman and Karén Simonyan to start the division in March 2024. Langford’s summary of the philosophy there: “scaling, and aggressively not research.” In that setting, he says, the main work becomes getting and refining data, which is far from what his group did. He adds, without elaboration, that there was “plenty of evidence that they just did not really want us.”

He also states the other side fairly. Bigger models do predict better, and from a business perspective the returns to scale are far more predictable than returns to research. A company committing capital at this level can reasonably prefer the bet it can forecast.

His rebuttal is an efficiency calculation. He says ByteDance experiments suggest a variant of the Full-Bandwidth Transformer might match a model twice its size, and asks what halving training and inference cost is worth when industry investment approaches a trillion dollars. Those experiments are not cited in the post, so the 2x figure cannot be checked from it. His own paper makes a narrower claim. It feeds the previous top-layer hidden state back in as input at each decoding step, trains 1B-parameter models on up to 400B tokens, and reports that they match or approach standard transformers trained with roughly 1.5x more tokens. That is a real result at small scale, and it is well short of proof at the frontier. Whether such gains hold at frontier sizes is the actual empirical question behind the dispute.

What it says about Microsoft Research

The most pointed passage concerns why he did not simply return to MSR. Langford gives two reasons. He could not get ideas into use, or even seriously considered, while inside Microsoft AI, so he sees little chance from further away. And he says the implicit bargain of the lab has broken: “we did good work and were non-consensually reorged away.” He would not be surprised to see that continue unless MSR’s leadership organizes to respond.

For a lab that has long sold itself on the freedom to do long-horizon work, that is a damaging charge from someone who spent years there. A commenter on the post, anonymous and unverifiable, says they are in another organization moved into Microsoft AI and feel the same way. That is an anecdote, but it suggests one question worth asking: whether AI Frontiers was an isolated case or part of a broader consolidation.

The practical consequences so far are modest but concrete. Langford says there will be no internships with the group’s researchers in the coming year, which closes a pipeline for PhD students who had worked with them.

What happens to AutoGen and the rest

Langford does not say what becomes of AutoGen, Fara or the monorepo. On AutoGen, the repository itself gives a partial answer that predates his post. Its README now states that the project is archived and no longer maintained, with no bug fixes, security fixes or support, and directs new users to Microsoft Agent Framework. The README commit making that explicit is dated October 16, following a maintenance-mode notice in April. GitHub’s own archived flag was still off when we checked on November 9, and the repository, with about 61,700 stars, last saw a push on October 28.

Anyone still running it should take the security language seriously. AutoGen Studio has already been the target of the AutoJack exploit chain, and archived agent frameworks are exactly where unpatched flaws accumulate. Microsoft has published a migration guide.

What to watch

The first thing to look for is where the group’s people land. Langford says many others value research, and a founding team of this caliber will be recruited quickly; a startup founder was already pitching for hires in the comments on his post. The second is whether Microsoft responds or names a successor effort for agents research and Fara. The third is technical: independent replication of full-bandwidth results at larger scale would give the research-versus-scale argument actual evidence. If the efficiency gains hold, Microsoft will have let the people behind them walk out. If they do not, the scaling camp’s preference for predictable returns will look justified.