Microsoft's AI Frontiers research group is over, says co-founder John Langford

AI Frontiers was a sub-organization inside Microsoft Research, the company's long-running research arm. It had about 50 people.

Langford says he and several like-minded colleagues created it in 2023, after ChatGPT was released. Their view was that research needed to be more organized to succeed than was typical at Microsoft Research.

The reason is simple. Other research groups were putting so much support and focus behind their work that one person with a clever idea would likely arrive late. One person could not move fast enough, and could not build baselines strong enough to show the idea was worth pursuing. A baseline is the standard result you compare a new idea against.

The big picture: AI Frontiers was a bet that modern AI research is a team sport.

Langford describes two kinds of work.

Much of the group's effort went into making AI agents work. An AI agent is an AI program that can take steps on its own to finish a task. Two examples he names:

• AutoGen: an orchestration framework, meaning software that coordinates several AI agents.
• Fara: a computer-use agent, which is an agent that operates a computer the way a person would.

In New York, a few people worked on new basic abilities for models:

• Next Latent: a training goal that Langford says enables extrapolation, which means making good guesses beyond the examples a model has seen.
• Dion3: an improved version of an optimizer called Muon. An optimizer is the part of training that adjusts a model as it learns.
• Full Bandwidth Transformer: a design that makes models work much more powerfully through recursion. Recursion here means reusing the same layers again.

Langford says that, given the group's small size and what it could do before, he had the general impression it was succeeding at pioneering things.

The group also tried to share. It decided to build a monorepo, one big shared code folder, that was open by default to everyone in Microsoft Research.

Inside was a pretraining codebase for edge scale. Pretraining is the first, biggest stage of building a model. Edge scale means models small enough to run on devices rather than in big data centers. Langford says he talked about it in his keynote at the COLM conference last month.

He calls the detail minor, but says building a well-formed model pipeline takes substantial effort. He knows of at least four efforts to build such pipelines, and three of them are not in repositories that are public inside Microsoft Research.

In real life Picture four teams in one company each building their own factory to make the same product. Sharing one factory sounds simpler.

In early September, the group was reorganized into Microsoft AI. Langford says that on the surface this could make sense. Putting two AI organizations together creates chances to focus effort, and each was ahead of the other in places. Microsoft AI is much larger.

In practice, he says, there were substantial issues. He describes the philosophy of the leadership there, which came from Inflection, as incompatible with the group's. The core idea is scaling, "and aggressively not research."

Scaling means making models bigger and feeding them more data and computing power. Langford grants that this has a logic: bigger models do make better predictions, and from a business view the returns to scale are far more predictable than the returns to research.

The catch: in a scaling-only setting, the main work is getting, making and refining data. Langford says that is "quite a step away" from what his group worked on. He also says there was plenty of evidence that the new leadership did not really want them.

Microsoft AI is led by Mustafa Suleyman. In March 2024, Microsoft paid a licensing fee of about $650 million to the startup Inflection AI and brought in its co-founders, including Suleyman and Karen Simonyan, plus roughly 70 employees, according to a report on the dispute by Winzheng.

That is where the "ex-Inflection leadership" Langford mentions comes from. Langford's post does not describe the deal.

Langford calls the scaling-only focus both misdirected and not fun. He says it is an understandable failure mode: people often give up on a tradeoff between two goals, like innovation and size, by choosing one and dropping the other.

He argues that returns to research are real. His example is the Full Bandwidth Transformer. He says experiments at ByteDance suggest one variant might be comparable to a model of twice the size.

Then he asks his readers to do the math. How much is it worth to halve training and inference costs, given that the industry has investments approaching a trillion dollars? Inference means running a trained model to get an answer.

These are Langford's own claims about results at another company. We could not check the ByteDance experiments ourselves. The post itself has no outside assessment of the group's work, as this summary points out.

Langford says he talked with people about returning to Microsoft Research, and found two serious obstacles.

• Research that goes nowhere: he is not comfortable at a company where his research has no path to real use. He says he could not get innovation into use, or even seriously considered, in Microsoft AI, so the prospects from Microsoft Research look small.
• A broken deal: he fears the basic deal at Microsoft Research, "do good work and good things happen," has broken down. As he sees it, the group did good work and was "non-consensually reorged away."

He also says he would not be surprised to see that kind of thing continue until leadership in Microsoft Research develops a will to organize effectively for the moment.

He ends on a hopeful note. He can imagine a future Microsoft that is attractive for research again, if the people in control develop a little more wisdom. He says he is fortunate that many other people in the world value research and innovation.

This is one person's account, and the post carries no statement from Microsoft. The company's reasons for the reorganization, and what happens to AutoGen, Fara and the other projects, are not covered in it.

Langford says every other founding leader of AI Frontiers seems to have come to a similar conclusion. He does not name them, and he does not say where they are going.

What's next: the one concrete change he states is small and human. There will be no internships with the group in the coming year.

• Products: the post describes a research group and its people. It says nothing about changes to Copilot, Windows or any other product you use.
• Open projects: if you use tools such as AutoGen, watch the project's own announcements for what comes next. The post does not say.
• The debate: the argument is whether better AI comes from bigger models, from new ideas, or both. Expect to hear it often.