Nathan Lambert is leaving the Allen Institute for AI (Ai2), where he led post-training work on the OLMo models and co-authored Tülu 3, a project that beat Meta's Llama 3 post-training using Llama's own base model. He arrived in October 2023 after a viral HuggingFace blog post on RLHF, a cold pivot from MEMS engineering, and a rejected request to be advised by Sergey Levine. He is leaving to work on coordinating the open AI ecosystem.
The most useful part of this piece is not the farewell. It is Lambert's unsentimental accounting of how influence gets built outside the frontier. OLMo was never the best-performing model in its size class, yet it shaped how open-weight research gets done. Lambert credits ruthless prioritization, zero medium-sized projects that went nowhere, and an institutional culture at Ai2 that sits between academia and industry in a way few organizations manage. He also flags a structural warning: AI needs independent voices precisely as it becomes more geopolitical and economically central.
What comes next is unspecified beyond a stated goal of making the open ecosystem better coordinated. The piece is worth reading in full for the career arc, the inside account of how Tülu 3 came together, and Lambert's argument that finding the right environment is harder than doing the work itself. Contact: nathan@natolambert.com.
[READ ORIGINAL →]