The End of Microsoft Research's AI Frontiers Group
After being folded into Microsoft AI, the AI Frontiers group's founding leadership left. The team worked on Autogen, Fara, and new training objectives including the Full Bandwidth Transformer.
Formation Inside MSR
After ChatGPT’s release in 2023, the author concluded that research at Microsoft Research needed more organization to succeed. A single person pursuing a clever idea was likely to arrive late and unconvincing, because baselines assembled by one person were too weak. Several colleagues agreed, and they formed the AI Frontiers sub-organization within MSR.
AI Frontiers tried to organize coherent research efforts. The approach was neither top-down nor bottom-up, but somewhere in the messy middle. Disagreements existed, and compute availability was often a severe bottleneck. Despite its small size of roughly 50 people, the group produced work the author considered pioneering. Part of the effort focused on making agents work, including the Autogen orchestration framework and the Fara computer use agent. A group in New York explored new fundamental capabilities: the Next Latent objective for extrapolation, Dion3 as an optimized Muon, and the Full Bandwidth Transformer for more powerful recursion.
Reorganization into Microsoft AI
In early September, AI Frontiers was reorganized into Microsoft AI. On the surface, combining two AI organizations could concentrate effort, and each side had areas where it was ahead. In practice, the philosophy of MAI’s ex-Inflection leadership proved incompatible. Their core philosophy is scaling, and aggressively not research. In a scaling-only environment, the primary effort goes into getting, making, and refining data—far from what AI Frontiers had been working on. The author also saw plenty of evidence that MAI did not really want the group.
The author describes the scaling-only focus as misdirected and not fun. Scaling is straightforward, and it is common to adopt one ideal completely while dropping the other. But returns to research are real. Bytedance experiments suggest a variant of the Full Bandwidth Transformer might be comparable to a model of twice the size, implying significant training and inference savings. With industry investments approaching a trillion dollars, halving those costs would be substantial.
Departure and the Question of MSR
The author situated his team as best he could and left. Every other founding AI Frontiers leadership team member appears to have reached a similar conclusion. Returning to MSR was considered but faced two obstacles. He was not comfortable at a company where his research had no path to real use; at MAI, innovation could not get into use or even be seriously considered. He also feared a breakdown of the basic deal in MSR: do good work and good things happen. As far as he could tell, they did good work and were non-consensually reorganized away.
AI Frontiers is over. There will be no internships with that team in the coming year. The author offers best wishes to those left behind and sees a possible future where Microsoft becomes attractive for research again, if those in control develop more wisdom.