Nathan Lambert
AI Researcher & Author, Interconnects
About
Nathan Lambert is a machine learning researcher known for his work on post-training and reinforcement learning from human feedback (RLHF). He earned his PhD at UC Berkeley working on model-based reinforcement learning for robotics, helped build the RLHF research team at Hugging Face, and later led post-training at the Allen Institute for AI (Ai2), where he shaped fully open model families like OLMo and Tülu. He is the author of the reference textbook on RLHF and writes Interconnects, a widely read technical newsletter on AI models and research. In 2026 he departed Ai2 to found a new AI lab, and remains one of the most visible advocates for open-source AI development in the United States.
Key Contributions
- Led post-training at the Allen Institute for AI (Ai2), shaping fully open model families like OLMo and Tülu with open weights, data, and training code
- Wrote the reference textbook on reinforcement learning from human feedback (RLHF), distilling the technique behind modern chat models
- Helped build the RLHF research team at Hugging Face, contributing to open tooling like TRL and early open aligned models like Zephyr
- Writes Interconnects, a technical newsletter on AI models and research that draws millions of views annually
- A leading advocate for open-source AI in the US; his 2026 visit to Chinese AI labs offered a rare firsthand American account of China's open-model ecosystem
Videos & Interviews
对话Nathan Lambert：美国知名AI研究者眼中的中国AI
Silicon Valley 101 (硅谷101) interview with Nathan Lambert about his 2026 visit to China's AI ecosystem, where he met teams at Alibaba, Moonshot AI, Zhipu, Tsinghua, and others — his firsthand impressions of Chinese researchers' pragmatism and the influence of DeepSeek on the open-model landscape.
View Details
State of AI in 2026: LLMs, Coding, Scaling Laws, China, Agents, GPUs, AGI | Lex Fridman Podcast #490
A four-hour conversation with Nathan Lambert and Sebastian Raschka on the state of AI in 2026 — scaling laws, open versus closed models, the training pipeline (pre-, mid-, post-training), coding agents and dev tooling, US-China competition, GPUs, and AGI timelines.
View DetailsConnections
Dario Amodei
In contrastCEO & Co-founder, Anthropic
The July 2026 open-weights fight placed them on opposite sides of a technical question with policy consequences. Amodei, whose lab publishes no weights, called safe open models a public good while asking for a crackdown on industrial-scale distillation; Lambert, who had spent years releasing OLMo's weights, data, and training code, argued publicly that distillation matters less each year as post-training shifts toward reinforcement learning. What looks like a disagreement about the value of openness is partly a disagreement about what copying a model can still achieve.
techcrunch.com
Jensen Huang
KindredFounder, President & CEO, NVIDIA
Both argue that open weights serve American interests, and the reasons could hardly be more different. Huang published his 2026 letter as the vendor whose business improves when models become a commodity and compute becomes the scarce thing; Lambert argues from the researcher's side, where a model you cannot inspect, retrain, or reproduce is not really an object of study. The pairing is worth holding because it shows openness is not one position but a place where a commercial interest and a scientific one happen to point the same way.
