The End of AI Frontiers

John Langford explains that AI Frontiers was formed at Microsoft Research after ChatGPT’s launch to organize research more effectively than isolated projects could. Its roughly 50 researchers combined coordinated efforts on agents, including AutoGen and the Fara computer-use agent, with work on core model capabilities such as the Next Latent objective, Dion3, and the Full Bandwidth Transformer. The group also built shared infrastructure, including a model-training codebase, so its work could benefit researchers across Microsoft. Langford says these efforts produced pioneering results despite limited compute and the challenges of aligning a research organization.

After AI Frontiers was moved into Microsoft AI, Langford says its research-led approach clashed with leadership focused on scaling models through data and larger training runs. He accepts that scaling can reliably improve predictions and business returns, but argues that research can deliver substantial efficiency gains: he cites experiments suggesting a Full Bandwidth Transformer variant could match a model twice its size, reducing training and inference costs. Concluding that the new organization had little appetite for research and that returning to Microsoft Research offered no dependable route from research to practical use, he left. Other founding leaders reached similar conclusions, and AI Frontiers ended, including its internship program. Langford says Microsoft could again become attractive to researchers if its leadership commits to organizing effectively and putting innovation into use.