Research leader departs Ai2 warning of closed-door AI development

Nathan Lambert has left the Allen Institute for AI (Ai2), where he played a central role in developing open-source language models including Olmo and Tülu. Before departing, Lambert reflected on his tenure and broader concerns about the direction of AI research. During his two years at Ai2, Lambert made influential contributions: creating RewardBench for reward model evaluation, enabling the release of Tülu 2 as the first public 70B-scale model to effectively implement direct preference optimization, and leading post-training initiatives including Tülu 3, which introduced the term Reinforcement Learning with Verifiable Rewards (RLVR). Lambert emphasizes that Ai2 operates in a rare space between academic and industrial research, enabling transparent model development. He warns that accelerating AI advances are increasingly confined to private corporations, concentrating power in few hands. He argues that independent researchers and open-source ecosystems are critical to ensuring AI technology benefits everyone and remains subject to diverse perspectives.