Sunday, 8 November 2026
Rīga TV

World and Latvian news in one place

TechnologyPublished: 8 November 2026 at 06:13

Microsoft Research's AI Frontiers group comes to an end

After being reorganized into Microsoft AI, the head of the AI Frontiers group has left the company and declared the group over. He criticizes a scaling-only approach and argues that research delivers real returns.

The AI Frontiers group inside Microsoft Research, the company's research division, no longer exists, according to a blog post on hunch.net by its founder and leader. He says he left Microsoft after the group was reorganized in early September, and that the other founding members of its leadership team reached the same decision. The post exposes a clash between research-driven and scaling-driven approaches to building large AI models.

How the group started

AI Frontiers was set up in 2023, after the release of ChatGPT. The author felt research needed to be more organized than was typical in Microsoft Research, because a single researcher often could not move fast enough or build baselines strong enough to show an idea's worth. The group had roughly 50 people. Much of its work focused on making AI agents function, such as the Autogen orchestration framework and the Fara computer-use agent. The New York team worked on fundamental capabilities, including the Next Latent objective, Dion3 (an improved Muon optimizer) and the Full Bandwidth Transformer architecture.

Reorganization and departure

In early September the group was folded into Microsoft AI (MAI), which is considerably larger. According to the author, the philosophy of MAI's ex-Inflection leadership is scaling rather than research, with the main effort going into obtaining and refining data. He considers this approach misdirected, and cites Bytedance experiments suggesting a variant of the Full Bandwidth Transformer may be comparable to a model twice the size, which would mean substantial savings in training and inference costs.

The author explains why he did not return to Microsoft Research: he doubts research there has a path to real use, and he believes the basic deal of "do good work and good things happen" has broken down. No internships with the group will be offered in the coming year.

Comments

0/1500

Comments are automatically moderated. No hate, threats, personal data or spam.

Loading comments…

More in this category