John Langford quit Microsoft on Nov. 6. He ran AI Frontiers, a 50-person research group inside Microsoft Research1. That group moved into the company's scaling-focused Microsoft AI unit two months earlier, on Sept. 142. Langford says the new bosses only wanted bigger models, and treated smarter ones as someone else's job.
Three Years, One Reorg, One Resignation
Langford helped start AI Frontiers in 2023, after ChatGPT's debut made rivals move fast1. He wanted a group that could organize research better than Microsoft Research typically did. The team built AutoGen, a framework for AI agents, and Fara, a tool that operates a computer on its own. Compute access stayed tight the whole time, Langford says.
Langford's name carries weight in the field. He held the title of partner research manager at Microsoft Research3.
Vowpal Wabbit, an open-source machine learning system, was his creation. He also led the ICML conference as its president, from 2019 to 2021. His next stop remains unannounced.
The group also built a shared codebase, open to the rest of Microsoft Research. It included a pretraining pipeline Langford described at a conference in October1. He counted four separate teams inside Microsoft building similar pipelines at the same time.
Microsoft moved the group into Microsoft AI on Sept. 14, 20262. Mustafa Suleyman runs that unit. He built its team around people he brought from Inflection, the startup he co-founded before Microsoft hired him in 2024.
Langford says the fit failed from the start. He left Nov. 6, two months after the reorg1. Every other founding leader of AI Frontiers reached the same conclusion, he wrote. Microsoft will skip internships through the group next year, he added.
What Scaling Crowded Out
Langford's team built something called the Full Bandwidth Transformer. It changes how a model reuses its own earlier work, a method called recursion. Bytedance tested a version and found it could match a model twice its size1.
Bytedance's own labs ran that comparison test, independent of Microsoft. An outside test carries more weight than a company's grade of its own work.
A standard model indexed at 100 needs an FBT variant scored at just 50 to match it, by that same test1. His New York group also built two related tools, the Next Latent objective and an optimizer called Dion3.
The Research Microsoft Shelved Cut Model Size by 50 Percent
Relative model size for the same score: standard scaling versus Langford's FBT variant, per Bytedance's tests. Lower is better.
The tall column is a standard-size model, indexed to 100. The short column is the size an FBT variant needs to match it, per Bytedance's tests. Lower is better.
Bytedance ran this test on its own, keeping the comparison independent of the company whose research group it describes.
Source: Machine Learning (Theory) [1]. Chart by Ryan Elliott Dennis.
The numbers behind this chart
| Item | Value |
|---|---|
| Standard scaling | 100 |
| FBT variant | 50 |
The size gap matters because of the money at stake. That industry's own spending is approaching $1 trillion, Langford wrote1. Halving training costs on that scale is worth billions, a number Microsoft has kept to itself.
Langford put it plainly in his resignation essay. "How much is it worth to halve training and inference cost given that the industry has investments approaching a trillion dollars?" he asked1. Microsoft executives have stayed silent on that question in public.
John Langford@JohnCLangfordThe End of AI Frontiers https://t.co/rQOIpmthGr .
There's a bit of a train wreck here that's painful to express. Nevertheless, it's a significant event for me and quite a few other folks involved so I wanted to memorialize a bit.
Microsoft Calls It a Natural Fit
Microsoft's own description of the move reads differently. The AI Frontiers lab's page still frames the change as continuity, a bigger step in the same direction. "MAI is a natural fit for our work, given MAI's focus on humanist superintelligence and building AI systems focused on real problems and in service to people," the page says2. The same page still lists Ece Kamar as CVP and lab director, the title she held before the merger.
Suleyman has bigger goals for his unit than AI Frontiers' old mandate. "Our job is to make sure that when we look out to 2030 and beyond, we have the capacity not just to buy models from third parties, but to build the absolute frontier, the best models in the world," he told VentureBeat in June4. He also described the unit as "what we think of as a hill-climbing machine, an organization that can continuously improve, cycle after cycle." Five months later, one founder of the group he absorbed walked out.
Langford disputes Microsoft's framing entirely. "I am just not comfortable being at a company where the research I do has no path to real use," he wrote1. He also says Microsoft AI's leadership runs on a different premise than his lab did.
Microsoft lost another senior research leader five weeks earlier. Peter Lee, president of Microsoft Science, left in October after 16 years at the company5. He gave a different reason for going: an AI advisory role with Mercedes' Formula 1 team, confirmed independently by two outlets6.
What to Do
Recruiters should watch where Langford's old team lands. A 50-person group with expertise in efficient model training is now on the market, and rivals that value research will notice1.
Executives should ask their own research teams Langford's question. What would it cost to match a model twice the usual size through smarter methods alone? The researcher who built that case for Microsoft just resigned.
Boards should ask whether research-versus-product fights at their own companies get settled by a memo or by a real conversation first. Microsoft's reorg played out as a memo, and Langford's resignation is the bill for it. Readers inside Microsoft AI should ask where their own ideas sit on that same spectrum. Otherwise they may end up explaining a resignation in a blog post of their own.
By the numbers
- 50 people: the size of the AI Frontiers group Microsoft folded into Microsoft AI1.
- 14 years: how long Langford worked at Microsoft Research before he left3.
- Two months: the gap between the Sept. 14 reorg and Langford's Nov. 6 resignation12.
- Twice the size: the model a Full Bandwidth Transformer variant can match, per Bytedance's tests1.
- Close to $1 trillion: the industry's own compute spending, by Langford's count1.
- 16 years: Peter Lee's tenure before he left Microsoft Science in October5.
Microsoft's Research Lab Lasted 2 Months Inside Its Scaling Unit
AI Frontiers' life, from its 2023 founding to John Langford's Nov. 6, 2026 resignation
Each pin is a dated event: the lab's 2023 founding, its Sept. 14 move into Microsoft AI, and Langford's resignation two months later.
Sources: Machine Learning (Theory) [1]; Microsoft Research [2]. Chart by Ryan Elliott Dennis.
The numbers behind this chart
| Date | Event |
|---|---|
| 2023 | AI Frontiers founded. About 50 researchers, after ChatGPT's debut |
| Sept. 14, 2026 | Folded into Microsoft AI |
| Nov. 6, 2026 | Langford resigns |
What to watch
Watch where AI Frontiers' other researchers land in the coming months. Several hires at one rival would back up Langford's account. New research hires inside Microsoft AI's own ranks would complicate it. Microsoft executives have kept quiet about Langford's account so far, and that quiet is itself worth watching.
Sources
- John Langford, "The End of AI Frontiers," Machine Learning (Theory), Nov. 6, 2026, https://hunch.net/?p=13764006
- Microsoft, "AI Frontiers," Microsoft Research, updated Sept. 14, 2026, https://www.microsoft.com/en-us/research/lab/ai-frontiers/overview/
- Windows Latest Staff, "'No path to real use': a veteran Microsoft AI scientist quits, says research that could halve AI costs wasn't even seriously considered," Windows Latest, Nov. 14, 2026, https://www.windowslatest.com/2026/11/14/no-path-to-real-use-a-veteran-microsoft-ai-scientist-quits-says-research-that-could-halve-ai-costs-wasnt-even-seriously-considered/
- Michael Nunez, "Microsoft AI chief says company was 'set free' from OpenAI to pursue superintelligence," VentureBeat, June 5, 2026, https://venturebeat.com/technology/microsoft-ai-chief-says-company-was-set-free-from-openai-to-pursue-superintelligence
- Todd Bishop, "Longtime Microsoft research leader Peter Lee and former LinkedIn CEO Ryan Roslansky to depart," GeekWire, Oct. 3, 2026, https://www.geekwire.com/2026/microsoft-executives-peter-lee-president-of-microsoft-science-and-evp-ryan-roslansky-to-depart/
- Sharon Goldman, "Exclusive: Microsoft's departing science president plans to join Mercedes F1 team as AI advisor," Fortune, Oct. 12, 2026, https://fortune.com/2026/10/12/microsoft-peter-lee-departing-science-president-plans-to-join-mercedes-f1-team-as-ai-advisor/
