AI Frontiers Organization at Microsoft Research Ends After Reorg

AI Frontiers, a sub-organization within Microsoft Research, is ending, according to a post on the Machine Learning (Theory) blog. The writer says the group was reorganized into Microsoft AI in early September and that its research-focused approach conflicted with the scaling-focused philosophy of Microsoft AI's ex-Inflection leadership. They also say they could not find a viable path to return to Microsoft Research and have left.

How AI Frontiers was formed

After the 2023 ChatGPT release, the writer says research needed more organization than was typical within Microsoft Research. Several people agreed and created AI Frontiers, which had about 50 people. The group tried to organize coherent research efforts in a middle ground between top-down and bottom-up approaches, though there were disagreements and compute shortages.

Its work included making agents work, such as the Autogen orchestration framework and the Fara computer use agent. In New York City, some members worked on fundamental capabilities, including the Next Latent objective, Dion3, and the Full Bandwidth Transformer. AI Frontiers also developed a monorepo shared by default with all of Microsoft Research, including an edge scale pretraining codebase mentioned in a COLM keynote.

Reorg into Microsoft AI

In early September, AI Frontiers was reorganized into Microsoft AI. The writer says that on the surface combining two AI organizations could improve concentration of effort, but in practice there were substantial issues. Microsoft AI's ex-Inflection leadership had a core philosophy of scaling and, in the writer's words, aggressively not research.

The writer describes the scaling-only focus as misdirected and not fun, while saying returns to research are real and significant. They cite Bytedance experiments suggesting a variant of the Full Bandwidth Transformer might be comparable to a model of twice the size, implying savings in training and inference. They also say there was evidence the larger organization did not really want the AI Frontiers group.

Why not return to Microsoft Research

The writer says they could not get innovation into use, or even seriously considered, when in Microsoft AI, so prospects for success from Microsoft Research seemed small. They also say they fear a breakdown of the basic deal in Microsoft Research: do good work and good things happen. They say good work was done but the group was non-consensually reorganized away.

AI Frontiers is over, the writer says. There will not be internships with the group in the coming year, and they offer best wishes to those left behind. They add that Microsoft could become attractive for research again if people in control develop more wisdom.

Sources: hunch.net · Business Examiner