Should we look for new ideas or simply build ever bigger models? The debate has agitated artificial intelligence for years. He has just found a concrete illustration at Microsoft. THE November 6, 2026the researcher John Langford announced on his blog the end ofAI Frontiersthe group he helped create within Microsoft Research. His post, entitled “The End of AI Frontiers”, describes a team absorbed by the Microsoft AI division, then effectively dissolved, due to lack of room for research in a strategy focused on the size of models.
A group born after ChatGPT
John Langford, a renowned machine learning expert and longtime author of the blog hunch.net, says that after ChatGPT was released in 2023, there was a need to organize search differently. According to him, an isolated researcher, even with a good idea, arrived too late and without results convincing enough compared to the means deployed by other laboratories.
With several colleagues, he created AI Frontiersa suborganization of Microsoft Research of approximately 50 peopleresponsible for leading coordinated research efforts. The researcher acknowledges internal disagreements and often very limited access to computing power, but believes that the group has carried out pioneering work given its size.
New agents and architectures
Much of AI Frontiers’ work focused on AI agentsthese programs which chain actions to accomplish a task. The group has notably developed:
| Project | Description according to John Langford |
|---|---|
| Autogen | An agent orchestration tool, allowing several agents to collaborate |
| Farah | An agent capable of using a computer |
| Next Latent | A training objective allowing extrapolation |
| Dion3 | An optimized version of the Muon optimizer, used to train models |
| Full Bandwidth Transformer | An architecture that makes models significantly more powerful thanks to recursion |
The group had also created a common code repository, shared by default with all of Microsoft Research, containing in particular a complete model training chain. A seemingly small detail, the researcher emphasizes, but building such a chain represents a considerable effort: he claims to know of at least four parallel internal projects aimed at the same objective.
Attachment to Microsoft AI
Beginning September 2026AI Frontiers has been integrated into Microsoft AI. On paper, the reorganization could be defended: bringing together two organizations dedicated to AI makes it possible to concentrate efforts, each having a head start on certain subjects.
To understand the scope of this rapprochement, we must recall the history of Microsoft AI. In March 2024, Microsoft recruited Mustafa Suleyman and a large part of the teams from his start-up Inflection AI to create a division responsible for consumer products like Copilot. Microsoft Research, for its part, maintained a more open and long-term research tradition, with numerous scientific publications.
In practice, according to John Langford, philosophies have proven incompatible. He describes the direction of Microsoft AI, which emerged from Inflection, as guided by scalingthat is, increasing the size of models and data, and “definitely not” through research. In such an environment, he explains, most of the effort focuses on collecting and refining data, far from the work of his team. He adds that there were many signs that the new management was not really interested in welcoming the group.
These comments are the sole responsibility of the author. Microsoft did not react in the sources consulted, and the post only presents the point of view of one of the parties.
Research or size: a fundamental debate
The researcher acknowledges that the scaling strategy is not absurd. There is ample evidence that larger models make better predictions, and from a business perspective, the returns from an investment in size are much more predictable than those of research.
Two ways to advance AI
Scaling
- Bigger models give better predictions
- More predictable returns on investment
- Effort focused on data and computing power
- Sharply rising training and usage costs
Exploratory research
- New architectures and training methods
- Uncertain and more distant results
- Potential cost savings, according to John Langford
- Requires time, freedom and access to computing
However, he considers this exclusive orientation “ill-advised”. According to him, the gains from research are real and significant. He cites experiments carried out by the Chinese group ByteDance, which suggest that a variant of its Full Bandwidth Transformer architecture could be comparable to a model twice the sizewhich would allow significant savings on both training and use. “Do the math,” he writes, recalling that investments in the sector are approaching a trillion dollars: halving the costs would have considerable value.
The departure of the founders
John Langford says he found a place for his team members and then left the company. According to him, all the other founding members of the AI Frontiers leadership seem to have come to a similar conclusion. The group will not offer courses next year.
Why not go back to Microsoft Research? The researcher puts forward two reasons. First, he does not want to work in a company where his research has not been no chance of being used in products. Then, he believes that the laboratory’s basic pact, “do good work and good things happen”, is broken: his team, he says, did good work and yet was reorganized without his agreement. He concludes by hoping that a future Microsoft will once again become attractive for research.
A movement that goes beyond Microsoft
The situation described is not unique to Microsoft. Since the rise of large language models, several companies have brought their research laboratories closer to their product teams, to accelerate commercialization. This movement gives researchers more means of calculation, but reduces the space for more exploratory work, the impact of which is uncertain and distant.
Defenders of fundamental research point out that several major advances in AI, such as the Transformer architecture, published in 2017 by Google researchers, were born from exploratory work before becoming the basis of the entire industry. Proponents of scaling respond that recent progress has mainly been due to increases in computing power and data.
What this can change
For Microsoft, the departure of experienced researchers could benefit competitors or young companies, which are actively recruiting in this field. For the entire sector, John Langford’s testimony poses a central economic question: while the cost of training and using models becomes a major issue, do companies that focus everything on size not risk missing out on innovations capable of doing as much with less? The answer will be found in the next generations of models, and in the capacity of large language models to progress without seeing their costs explode.
Frequently asked questions
What was AI Frontiers at Microsoft?
AI Frontiers was a suborganization of Microsoft Research of approximately 50 people, established in 2023 after the release of ChatGPT. She was carrying out coordinated research on AI agents, with the Autogen tool and the Fara agent, and on new model architectures. It was attached to Microsoft AI in September 2026.
Why did John Langford leave Microsoft?
According to his November 6, 2026 post, Microsoft AI management favors increasing the size of models and data over research, a philosophy he considers incompatible with his work. He also believes that his research no longer had a chance of being used in products.
What is scaling in artificial intelligence?
Scaling involves improving models by increasing their size, the amount of data, and the computing power used to train them. This strategy has produced steady and predictable progress, but it significantly increases training and usage costs.
Who runs Microsoft AI?
Microsoft AI is led by Mustafa Suleyman, co-founder of DeepMind and then of the start-up Inflection AI, who arrived at Microsoft in March 2024 with a large part of the Inflection teams. The division notably develops the Copilot assistant and AI models designed in-house.
Sources
References consulted for the writing of this article. The addresses are provided for information purposes only and are not links.
- John Langford, The End of AI Frontiers, hunch.net, November 6, 2026hunch.net/?p=13764006
Report an error or inaccuracy in this article

