Nathan Lambert

Nathan Lambert is a machine learning researcher focused on large language model post-training and reinforcement learning from human feedback (RLHF), who worked on open models such as Zephyr, OLMo and Tülu at Hugging Face and Ai2, writes the Interconnects newsletter and an RLHF textbook, and co-founded the nonprofit Trillium Labs in 2026.

Contents22 sections

Key facts

2013–2017Studies at Cornell University

From 2013 to 2017 he studied electrical and computer engineering at Cornell University Cornell UniversityCornell University is a private research university in Ithaca, New York, and a member of the Ivy League. It is New York State's federal land-grant university, a distinction unique within the Ivy League. Ezra Cornell and Andrew Dickson White founded the university in 1865, and it opened in 1868. In addition to its Ithaca campus, Cornell operates Weill Cornell Medicine and Cornell Tech in New York City and a medical college in Qatar. It enrolled 26,793 students in fall 2024. Its president is Michael I. Kotlikoff.Open the full entry.⁠[1]

2017Begins a PhD at UC Berkeley

In fall 2017 he began a PhD at UC Berkeley University of California, BerkeleyThe University of California, Berkeley (UC Berkeley) is a campus of the University of California system; the University of California was founded in 1868 through a merger of two institutions and moved to Berkeley, California, in 1873.Open the full entry; he recalls that his requests to join the groups of Sergey Levine Sergey LevineSergey Levine is a computer scientist and associate professor of electrical engineering and computer sciences at UC Berkeley who works on deep reinforcement learning and robot learning, co-developed algorithms such as Soft Actor-Critic and Conservative Q-Learning, and in 2024 co-founded the robot foundation model company Physical Intelligence.Open the full entry and Pieter Abbeel were declined, and that his advisor Kristofer Pister later connected him with Roberto Calandra.⁠[2]

2019–2021Internships at Facebook AI and DeepMind

He interned at Facebook AI from 2019 to 2020 and remotely at DeepMind Google DeepMindGoogle DeepMind is Google’s artificial intelligence research organization, formed in April 2023 by merging DeepMind, which Google acquired in 2014, with the Brain team from Google Research, and led by CEO Demis Hassabis. It developed AlphaGo, AlphaFold, and the Gemini family of models, and Hassabis and John Jumper received the 2024 Nobel Prize in Chemistry for AlphaFold’s protein structure prediction work.Open the full entry in 2021.⁠[1]

March 2021Appears on the TalkRL Podcast

In March 2021, as a PhD candidate at UC Berkeley, he appeared on the TalkRL podcast to discuss model-based reinforcement learning, trajectory-based models and quadrotor control; the episode’s references include “Low Level Control of a Quadrotor with Deep Model-Based Reinforcement Learning”, which he co-authored with Roberto Calandra, Sergey Levine, Kristofer Pister and others.⁠[3]

2022Completes His Doctoral Thesis

In 2022 he completed his doctoral thesis, “Synergy of Prediction and Control in Model-based Reinforcement Learning” (Technical Report UCB/EECS-2022-65), advised by Kristofer Pister; it studies how a learned dynamics model interacts with decision making.⁠[4]

May 2022Joins Hugging Face

In May 2022 he joined Hugging Face; he recalls that Douwe Kiela recruited him to help build an “open-source DeepMind”.⁠[2]

December 2022Publishes an Illustrated Guide to RLHF

On December 9, 2022 he, Louis Castricato, Leandro von Werra and Alex Havrilla published “Illustrating Reinforcement Learning from Human Feedback (RLHF)” on the Hugging Face blog, followed by a lecture on December 13.⁠[5]

October 2023Co-Authors the Zephyr Paper

He is listed as third author of the Zephyr paper, “Zephyr: Direct Distillation of LM Alignment”, released in October 2023.⁠[6]

2023Joins Ai2

He recalls that Ai2 was training Tülu 2 70B when he joined, and that Chris Manning later credited him with “saving DPO”.⁠[2]

February–March 2024: Contributes to OLMo and Releases RewardBench

In February 2024 Dirk Groeneveld and colleagues released the open language model paper “OLMo: Accelerating the Science of Language Models”, of which he is a co-author.⁠[7]

That March he released the reward-model benchmark “RewardBench” as first author with Valentina Pyatkin, Tom Zick and others.⁠[8]

2024Projects He Worked on Win ACL and GeekWire Awards

In 2024 OLMo, on which he worked, won the ACL Best Theme Paper Award and GeekWire’s Innovation of the Year, and Dolma won the ACL Best Resource Paper Award.⁠[1]

November 2024Leads the Tülu 3 Release

On November 21, 2024 Ai2 released the Tülu 3 instruction-following model family together with its data, data mixes, recipes, code, infrastructure and evaluation framework, introducing reinforcement learning with verifiable rewards.⁠[9]

The technical report “Tulu 3: Pushing Frontiers in Open Language Model Post-Training”, released the next day, lists him as first author.⁠[10]

February 2, 2025First Appearance on the Lex Fridman Podcast

He appeared with SemiAnalysis SemiAnalysisSemiAnalysis is a semiconductor and AI research firm founded by Dylan Patel in 2020. It produces reporting, institutional data models, consulting and hardware testing, covering chip manufacturing, memory, datacenters, inference costs and AI economics.Open the full entry founder Dylan Patel Dylan PatelDylan Patel is the founder and CEO of SemiAnalysis, researching semiconductor supply chains, datacenters and AI economics. He started the business in 2020 and expanded individual industry writing into data models, consulting and technical testing, while discussing inference costs and hardware–software co-design.Open the full entry in episode 459 of the Lex Fridman Lex FridmanLex Fridman is a Soviet-born, US-raised computer scientist and podcast host who works as an MIT research scientist on human-AI interaction and autonomous driving. His Lex Fridman Podcast is known for hours-long conversations with guests including heads of state and technology leaders.Open the full entry Podcast, whose page describes him as a research scientist at Ai2 and author of the AI blog Interconnects; topics included DeepSeek-R1 and V3, training costs, GPU export controls, TSMC TSMCTSMC is a semiconductor manufacturer headquartered in Hsinchu, Taiwan, that pioneered the pure-play foundry model and manufactures chips for hundreds of customers worldwide.Open the full entry and Taiwan, AI megaclusters and open source.⁠[11]

August 2025Launches the ATOM Project

On August 4, 2025 he launched the American Truly Open Models (ATOM) Project to advocate building truly open models in the United States; signatures were collected from August 2025 through September 2026, with 395 researchers, builders and leaders signing on.⁠[12]

January 31, 2026Second Appearance on the Lex Fridman Podcast

He appeared with Sebastian Raschka Sebastian RaschkaSebastian Raschka is a machine learning and large language model research engineer and technical author, formerly an assistant professor of statistics at the University of Wisconsin–Madison and a research engineer at Lightning AI, whose books include Python Machine Learning and Build a Large Language Model (From Scratch).Open the full entry in episode 490, whose page calls him post-training lead at Ai2; topics included the US–China AI race, open versus closed models and post-training.⁠[13]

April 2026Publishes the ATOM Report

In April 2026 he and Florian Brand released “The ATOM Report: Measuring the Open Language Model Ecosystem”, surveying about 1,500 mainline open models and documenting that Chinese models overtook their U.S. counterparts in summer 2025 and widened the gap.⁠[14]

May 7, 2026Publishes Notes from China’s AI Labs

After visiting most of the leading AI labs in China, he published “Notes from inside China's AI labs” on May 7, 2026, arguing that Chinese labs are culturally and organizationally well suited to fast-following the frontier.⁠[15]

July 2026Manning Publishes “Reinforcement Learning from Human Feedback”

His textbook on RLHF and LLM post-training was published by Manning in print, ePub and liveBook editions in July 2026; the companion course followed in August, when Amazon AmazonAmazon is a publicly traded U.S. technology and retail company founded by Jeff Bezos in 1994, with its principal corporate offices in Seattle. Its businesses span e-commerce, logistics and fulfillment, digital advertising, membership and entertainment services, consumer electronics, and Amazon Web Services cloud computing.Open the full entry sales began, with a first reprint in October.⁠[16]

The publisher’s author bio says he has released models, datasets and training code through Ai2 and Hugging Face, including Zephyr-Beta, Tulu 2, OLMo, TRL and Open Instruct.⁠[17]

October 1, 2026Co-Founds Trillium Labs

On October 1, 2026 he and Tom Zick announced the nonprofit Trillium Labs to sustain open science in frontier AI, starting with post-training and releasing full recipes, data, code, evaluations and intermediate checkpoints; it began with support from Halcyon Futures and Schmidt Sciences.⁠[18]

Trillium Labs has offices in the San Francisco Bay Area and Cambridge, Massachusetts, and its advisors include Thomas Wolf Thomas WolfThomas Wolf is a co-founder and the chief science officer of Hugging Face who has helped drive open-source tools such as Transformers, the BigScience open research collaboration, and robotics projects. He worked in physics research and patent law before turning to machine learning, and his public work focuses on lowering the barriers to using AI models, data, and development tools.Open the full entry, Hanna Hajishirzi, Graham Neubig and Bryan Catanzaro.⁠[19]

As of October 2026: Executive Director of Trillium Labs and Author of Interconnects

As of October 2026 his personal website describes him as co-founder and executive director of Trillium Labs, and lists The Retort, the AI podcast he ran with Thomas Krendl Gilbert, as retired.⁠[20][21]

Interconnects, the newsletter he founded and runs, publishes one to three times a week with essays, model reviews, researcher interviews and surveys of the open-model ecosystem, and has passed 300 posts since he began writing weekly after ChatGPT’s release.⁠[22]