Nathan Lambert is a machine-learning researcher, founder of Interconnects and The ATOM Project, and author of an open textbook on reinforcement learning from human feedback. Formerly the post-training lead at the Allen Institute for AI, he helped develop Tülu 3 and the Olmo family of open language models, advancing the case for making model weights, training methods, datasets, and evaluations accessible to independent researchers.
Lambert studied electrical and computer engineering at Cornell and completed a doctorate at the University of California, Berkeley, where he worked on robotics, reinforcement learning, and autonomous systems. He subsequently became a research scientist at Hugging Face, helped develop its RLHF research program, and coauthored an influential introduction to training language models with human feedback.
At the Allen Institute for AI, he worked on Olmo and led post-training research for open language models. His account of entering AI research connects his progression from robotics to Hugging Face and Ai2 with a growing commitment to public technical writing and reproducible research.
He founded Interconnects to cover open models, post-training, and AI’s shifting economics, and wrote Reinforcement Learning from Human Feedback, an openly accessible textbook spanning instruction tuning, reward modeling, preference optimization, and reinforcement learning. Its public GitHub repository includes the book’s source, code, and reusable diagrams.
- Open post-training through Tülu 3. As first author of the Tülu 3 paper, Lambert helped release a complete post-training recipe combining supervised fine-tuning, preference optimization, and reinforcement learning with verifiable rewards. Public models, datasets, code, and evaluations make the resulting capabilities reproducible; verifiable rewards train models on tasks whose answers or instruction-following constraints can be checked directly.
- Skills, calibration, strategy, and abstraction. Lambert’s framework for reasoning models distinguishes basic capabilities such as coding, mathematics, and search from the harder requirements of allocating compute appropriately, abandoning failed approaches, and decomposing complex objectives. A system that retrieves obscure information brilliantly can still fail to organize a dependable research project.
- Planning as a trainable capability. He argues that reliable long-horizon execution requires expert-written plans, supervised fine-tuning, reinforcement learning on progressively harder tasks, and stable training infrastructure. Better planning can improve memory management, reduce repeated mistakes, coordinate parallel subtasks, and prevent expensive reasoning traces on simple questions.
- Alignment without sacrificing useful behavior. His writing on RLHF tradeoffs examines how imperfect reward signals and excessive optimization can make models more compliant while degrading useful capabilities.
