Takeaway: John Langford, a co-founder of Microsoft Research’s AI Frontiers group, published a post titled “The End of AI Frontiers” on his blog, hunch.net, dated November 6, 2026, and appearing with a November 7 stamp as well. He wrote that he and colleagues created the roughly 50-person group inside Microsoft Research in 2023, after ChatGPT’s release, because research needed more organization than MSR had typically provided. In early September 2026, he said, the group was reorganized into Microsoft AI, whose former Inflection leadership he described as committed to “scaling, and aggressively not research.” There was also, he wrote, “plenty of evidence that they just did not really want us.” He said he “situated my folks as best I could and left,” that every other founding member of the AI Frontiers leadership team “seems to have come to a similar conclusion,” that he will not return to Microsoft Research, and that there will not be internships with “the folks there” in the coming year. His closing line is unambiguous: “So AI Frontiers is over.” Microsoft has not issued a formal shutdown press release. Its AI Frontiers page still carries a September 14, 2026 update saying the lab “is now part of Microsoft AI” and will pursue “our same goal of frontier and cutting edge agentic and model research” there.
This is an organizational story about how one of Microsoft’s research groups was moved, how a founding leader described what happened next, and how Microsoft’s own page still frames the same move. It is not a patent story. It is distinct from this morning’s note on a Microsoft patent application that describes rewriting livestream speech and redrawing lips. It also sits beside, rather than inside, other research-pace pieces we have covered this week, including METR’s write-up of Anthropic’s AI research-and-development speed and Anthropic’s own post on agent memory and dreaming.
What Langford said AI Frontiers was built to do
Langford’s account begins in 2023. After ChatGPT’s release, he wrote, it was clear to him that research needed to be more organized to succeed than was typical inside Microsoft Research. Other groups, he said, were applying enough support and focus that a clever idea pursued by one person was likely to arrive “late and unconvincing,” because one person could not move fast enough and the baselines one person could assemble were not strong enough. “Several other folks thought similarly,” he wrote, “and we organized to create the AI Frontiers sub-organization within MSR.”
Inside the group, he said, the work of organizing coherent research sat “somewhere in the messy middle” between top-down and bottom-up direction. There were disagreements, and compute was “an often-severe bottleneck.” Even so, he wrote, he had the impression the organization was succeeding in doing pioneering things, “particularly given the small size of AIF (~50 people) and relative to what we could do previously.”
He named the work in two clusters. Much of AI Frontiers, he said, was about making agents work, “for example the Autogen orchestration framework and the Fara computer use agent.” In New York, “several of us were interested in developing new fundamental capabilities,” including the Next Latent objective, “which enables extrapolation”; Dion3, “which creates an optimized/improved Muon”; and the Full Bandwidth Transformer, “which efficiently makes models function substantially more powerfully via recursion.”
The group also tried to share infrastructure with the rest of Microsoft Research. AI Frontiers, he said, developed a monorepo that was shared by default with all of MSR. In that monorepo was “the proven-capable edge scale pretraining codebase” he said he had discussed in a COLM keynote the previous month. Building a well-formed model-creation pipeline, he wrote, takes substantial effort, and he was aware of at least four such efforts — the other three not in public-within-MSR repositories.
The September move into Microsoft AI
“In early September we were reorged into Microsoft AI,” Langford wrote. On the surface, he said, that could make sense: putting two AI organizations together creates opportunities to concentrate effort, and there were places where AI Frontiers was ahead of Microsoft AI and vice versa. Microsoft AI, he noted, is much larger.
The problem, in his telling, was philosophical. “In practice, there are substantial issues because the philosophy of the ex- Inflection leadership of MAI is incompatible. The core philosophy there is scaling, and aggressively not research.” He wrote that this might shock people who assume better models come from research, but that there is “plenty of evidence that bigger models are better models as far as the quality of the predictions they make,” and that, from a business perspective, “returns to scale are far more predictable than returns to research.” In a scaling-only environment, he said, the primary effort is on getting, making, and refining data — “quite a step away from what we were working on.” Then came the sharper line: “Aside from the incompatible focus, there was plenty of evidence that they just did not really want us.” He did not itemize that evidence.
I consider the scaling-only focus to be both misdirected and not fun. It is an understandable failure mode—scaling is straightforward and it’s common for people to give up on a tradeoff between multiple competing ideals (like innovation and size) by short-sightedly adopting one in totality and dropping the other. But returns to research are real and significant.
John Langford, “The End of AI Frontiers,” November 6, 2026
As an example of those returns, he pointed to ByteDance experiments that, in his reading, “suggest a variant of the Full Bandwidth Transformer might be comparable to a model of twice the size,” which would imply significant savings in training and inference for the same capability. “Do the math—how much is it worth to halve training and inference cost given that the industry has investments approaching a trillion dollars?” That comparison is Langford’s interpretation of those experiments, not a Microsoft product claim and not a figure we independently audited.
Why he left — and why he says he will not go back to MSR
The personal decision follows directly from that diagnosis. “So I situated my folks as best I could and left.” He wrote that he is fortunate there are many other people who see the value of research and innovation, which he called his “intrinsic center of interest.” He also wrote, with a hedge that readers should keep attached to the sentence, that “every other founding AIF leadership team member seems to have come to a similar conclusion.” He did not publish a list of those people, and he did not say how many of the roughly 50 researchers are staying at Microsoft.
He then asked the question many readers would ask: why not go back to Microsoft Research? He said he had “chatted with folks a bit” and could not get past two obstacles. The first was practical. “I am just not comfortable being at a company where the research I do has no path to real use.” He said he “could not get innovation into use (or even seriously considered)” while he was in Microsoft AI, “so the prospects for success from MSR seem quite small.” The second was cultural. He wrote that he was afraid of “a breakdown of the basic deal in MSR: do good work and good things happen.” As far as he could tell, “we did good work and were non-consensually reorged away.” He added that he would not be surprised to see that kind of thing continue “until MSR leadership develops a will to effectively organize to meet the moment.”
So AI Frontiers is over. I’m sorry to say that there will not be internships with the folks there in the coming year—I enjoyed working with interns on many occasions. And my best wishes for the folks I leave behind. It is not hard to imagine a Microsoft of the future which becomes attractive for research again if the people in control can develop a little bit more wisdom.
John Langford, “The End of AI Frontiers,” November 6, 2026
When he shared the post on X on November 6, EngineerPal quoted him calling the situation “a bit of a train wreck here that’s painful to express.” India Today (November 7) and The Digital Brief News (November 8) recapped the same blog post and drew the distinction we draw below: Langford has declared the group over; Microsoft has not published a matching shutdown notice.
How Microsoft’s page still describes the same move
Microsoft’s Research page for AI Frontiers remains live. An update dated September 14, 2026, at the top of the “Who We Are” section, says: “Microsoft Research AI Frontiers is now part of Microsoft AI.” The page says the lab has long focused on advancing frontier research in efficient language models and agentic systems, and on “reimagining existing computing through an agent-native lens.” It lists shipped work that overlaps Langford’s account and goes beyond it: AgentInstruct, the Phi family of models, the Fara portfolio of computer-use agents, “the Agent Framework AutoGen,” Magentic-One, MagenticUI, and Magentic Marketplace, plus work on agent red teaming, benchmarking, and safety research.
The page then describes the September change as a continuation, not an ending: “We’ll pursue our same goal of frontier and cutting edge agentic and model research at Microsoft AI (MAI) starting now.” It calls Microsoft AI “a natural fit” and closes the update with, “We’ll share more exciting news on this transition soon!” Under “Leadership,” the page still lists Ece Kamar as CVP and Lab Director of AI Frontiers. It does not mention Langford’s departure, a shutdown, or a halt to internships.
Those two documents can both be true in their own lanes. Langford is describing what happened to him, to the founding leadership as he understands it, and to the internship pipeline he used to run. Microsoft’s page is describing an organizational transfer and a promise to keep pursuing the same research goals. A transfer is not automatically a closure, and a founder saying “AI Frontiers is over” is not automatically a company press release. As of this writing, we have not seen Microsoft publish a statement that uses shutdown language or comments on Langford’s post. Readers who depend on AutoGen, Fara, or other listed projects should watch Microsoft’s later updates rather than treat a personal blog post as a product-sunsetting notice — a contrast with explicit engineering changes such as GitHub Actions default-blocking pull_request_target.
The argument underneath the organogram
Langford’s post is also an argument about how AI progress happens. He does not deny that larger models often make better predictions, or that scale can look more predictable to a finance team than a research bet. His claim is that treating scale as the only lever is a short-sighted way to resolve a tradeoff. The Full Bandwidth Transformer example is there to make that concrete: if a method really did match a model twice the size, the implied saving would be large in an industry he described as approaching a trillion dollars of investment. That debate is bigger than one subgroup — a tension also visible in third-party measurements of how much AI is accelerating AI research itself — but it does not tell you whether Microsoft will later absorb, rename, or continue particular AI Frontiers projects.
What this is — and what it is not
| Claim | Status | Notes |
|---|---|---|
| Langford declared “AI Frontiers is over” | Stated by Langford, Nov 6–7 blog post | Not a Microsoft press release |
| AI Frontiers was created inside Microsoft Research in 2023, after ChatGPT, and had about 50 people | Stated by Langford | He described himself as one of several organizers, not as the sole director |
| Work included AutoGen, Fara, Next Latent, Dion3, and the Full Bandwidth Transformer | Named by Langford | Microsoft’s page also lists AutoGen, Fara, plus Phi, AgentInstruct, and Magentic projects |
| The group shared a monorepo, including an edge-scale pretraining codebase discussed at COLM | Stated by Langford | He said he knew of at least four model-pipeline efforts |
| The group was reorganized into Microsoft AI in early September 2026 | Stated by Langford; consistent with Microsoft’s Sept 14 page update | The page says the lab “is now part of Microsoft AI” |
| Microsoft AI’s ex-Inflection leadership philosophy is “scaling, and aggressively not research” | Langford’s characterization | Microsoft has not published a reply we have seen |
| There was “plenty of evidence that they just did not really want us” | Langford’s assessment | He did not list the evidence |
| Langford situated his people as best he could and left | Stated by Langford | He did not say how many of the ~50 remain |
| Every other founding leadership member reached a similar conclusion | Langford’s impression (“seems to have”) | Not a published roster |
| He will not return to Microsoft Research | Stated by Langford | Reasons: no path to real use; a breakdown of “do good work and good things happen” |
| No AI Frontiers internships in the coming year | Stated by Langford | His announcement, not a Microsoft careers bulletin we have seen |
| Microsoft issued a formal shutdown press release | No | The live Research page still frames a transition and promises more news |
| A ByteDance variant of the Full Bandwidth Transformer matched a model twice as large | Langford’s reading of ByteDance experiments | Not an independent audit by us, and not a Microsoft product claim |
| This is the same story as Microsoft’s livestream-speech patent application | No | That note is about a published patent filing, not a research-group reorg |
FAQ
Did Microsoft announce that AI Frontiers has shut down?
No. Langford wrote that “AI Frontiers is over.” Microsoft’s Research page, in a September 14, 2026 update that was still live when we checked, says the lab is now part of Microsoft AI and will pursue the same research goals there. We have not seen a Microsoft press release that uses shutdown language or that comments on Langford’s post.
What did John Langford say happened?
He said AI Frontiers, a roughly 50-person group created inside Microsoft Research in 2023, was reorganized into Microsoft AI in early September 2026. He described Microsoft AI’s former Inflection leadership as focused on “scaling, and aggressively not research,” said there was evidence the group was not wanted, placed his people as best he could, and left. He said he would not return to Microsoft Research and that the founding leadership team appeared to have reached a similar conclusion.
What did AI Frontiers work on?
In Langford’s post: AutoGen, the Fara computer-use agent, the Next Latent objective, Dion3, and the Full Bandwidth Transformer, plus a shared monorepo used across Microsoft Research. Microsoft’s own page lists several of those efforts and adds others, including the Phi models, AgentInstruct, Magentic-One, MagenticUI, and Magentic Marketplace.
Why does Langford say he will not return to Microsoft Research?
Two reasons, in his words. He is not comfortable being at a company where the research he does has “no path to real use,” and he could not get innovation used or even seriously considered inside Microsoft AI, which made the prospects from MSR look small. He also said the basic MSR deal — “do good work and good things happen” — had broken down when the group was “non-consensually reorged away.”
Will there be AI Frontiers internships next year?
Langford said there will not be internships “with the folks there in the coming year.” That is his statement. Microsoft’s AI Frontiers page does not address the internship cycle.
How does Microsoft describe the September change?
As a transition, not an ending. The September 14 update says Microsoft Research AI Frontiers is now part of Microsoft AI, that the team will pursue the same goal of frontier agentic and model research there, and that more news on the transition would come soon. The page still lists Ece Kamar as lab director.
Is this the same story as Microsoft’s livestream-speech patent?
No. This morning’s article is about U.S. patent application 2026/0341309 A1, which describes using AI to rewrite livestream speech and redraw a speaker’s lips. This article is about what a founding leader of AI Frontiers says happened after that group was moved into Microsoft AI, and about how Microsoft’s public page still describes the move.
Sources: John Langford, “The End of AI Frontiers,” hunch.net (dated 11/6/2026, with an 11/7/2026 stamp); Microsoft Research, AI Frontiers lab page (September 14, 2026 update, checked November 9, 2026); India Today, November 7, 2026; EngineerPal, November 8, 2026; The Digital Brief News, November 8, 2026. Direct quotations are from Langford’s post or from Microsoft’s page. Langford’s characterizations of Microsoft AI’s philosophy and of how wanted the group was are his; Microsoft has not published a formal response that we have seen.