John Langford Microsoft exit: why his AI lab lost to scaling

John Langford is a machine learning researcher who came to Microsoft Research in May 2012, where he was one of the people who started its New York City lab. Before that he worked at Yahoo! Research and IBM. At Microsoft he rose to the rank of Partner Research Manager.

His name is attached to tools and ideas researchers use every day. Vowpal Wabbit, a free and open machine learning toolkit, was built mainly by him. He also gave contextual bandits their name, a technique that helps apps and websites decide what to show each person. He also served as president of ICML, one of the largest machine learning conferences, from 2019 to 2021.

He explained his exit in a post on his personal blog titled “The End of AI Frontiers.” On X, where he shared it, he called the situation “a bit of a train wreck,” and that post has passed 330,000 views.

AI Frontiers was moved into Microsoft AI in early September 2026, and Langford says that move ended the lab in practice. He and several colleagues set the group up inside Microsoft Research in 2023, after ChatGPT arrived. Their reasoning: a single researcher chasing one good idea could no longer move fast enough, or build strong enough comparisons, to keep up with the large, organized teams at other companies.

The lab grew to about 50 people. Much of its effort went into making AI agents work, including the AutoGen framework and Fara, a model that can operate a computer. A New York group also worked on new basic building blocks for models, and the lab shared its model-building code with the rest of Microsoft Research by default.

Microsoft AI is far bigger. Satya Nadella set it up in March 2024, bringing in Mustafa Suleyman and Karén Simonyan, who had co-founded the startup Inflection, to run Copilot and Microsoft’s consumer AI products. Langford writes that merging the two groups made sense on paper. In practice, he says, the former Inflection leaders who run Microsoft AI follow a philosophy built on scaling (making models bigger and feeding them more data) and are firmly against research. He adds that there was plenty of evidence the division simply did not want his team.

He is fair to the other side. He accepts that bigger models do make better predictions, and that for a business the payoff from scaling is easier to forecast than the payoff from research. His objection is that a scaling-only shop spends most of its energy on gathering and cleaning data, which is a long way from what his group did. He calls that focus misdirected, and not fun.

The clearest example is the Full Bandwidth Transformer, a model design Langford published in August with seven co-authors. A normal language model works out a rich internal state at each step, then squeezes it into a single token (a word or piece of a word) and throws the rest away. His design feeds that fuller state back into the model together with the next token.

In the paper’s tests, 1-billion-parameter models, fed as many as 400 billion tokens during training, equaled or nearly equaled standard models trained on about 1.5 times as much data, while the extra cost for each token was negligible.

Then a rival ran the idea at bigger scale. On October 25, ByteDance, the company behind TikTok, published a method called Recycler that reuses hidden states in a similar way. Langford said on X that Recycler is essentially his design apart from minor details, and that ByteDance’s experiments were larger and more thorough than his.

Based on those results, he writes that a variant of his design might perform like a model twice its size. With industry investment approaching a trillion dollars, he invites readers to do the math on what halving training and running costs would be worth.

Two caveats matter. Halving costs is Langford’s own projection, not a measured saving, and matching a bigger model on tests does not automatically cut a real bill in half. He also does not say Microsoft AI turned down this particular paper. What he says is that while he was in the division, he could not get new ideas into use, or even seriously considered.

If you build with AutoGen, the practical news is that Microsoft has stopped maintaining it. The project page now marks it as archived and says Microsoft will offer no support, no new features and no fixes for bugs or security holes. The code stays online for reference, but Microsoft warns that unpatched flaws in it and its dependencies may linger, and it still advises against production use.

That notice went up in mid-October, a few weeks before Langford published his farewell post in early November. Microsoft now points new users to Microsoft Agent Framework, which it calls the enterprise-ready successor to AutoGen, and offers a migration guide for existing projects. Agent Framework 1.0 is described as production-ready, with stable interfaces and a commitment to long-term support.

For teams still running AutoGen, the archive notice spells out the risk in plain terms: without security fixes, any project handling sensitive data needs a careful review, and Microsoft’s own recommended path is to move to the successor framework.

Microsoft has not answered Langford publicly. The AI Frontiers web page, last changed on September 14, describes the group as folded into Microsoft AI while continuing frontier work. Microsoft AI has also said it is building a superintelligence lab, an announcement that came with the launch of seven MAI models at Build 2026.

There is an awkward contrast here. Microsoft AI has promoted cheaper models of its own, including a coding model it describes as better and faster at a quarter of the cost. So lower costs clearly matter to the division. Langford’s complaint is that it showed no interest when the savings came from his kind of research.

The departure also lands in the middle of a wider shake-up. Microsoft is offering buyouts to around 60,000 employees, while Nadella talks about a future with 20 million AI agents doing work. Langford says every other founding leader of AI Frontiers seems to have reached the same conclusion he did, and that the group will not take interns next year.

Why not just go back to Microsoft Research? He gave two reasons. He is not comfortable at a company where his work has no route into real products, and he believes the lab’s basic bargain, that good work leads to good things, broke when his team was reorganized without its consent.