Nathan Lambert is a machine learning researcher and co-founder and executive director of Trillium Labs, a nonprofit supporting open frontier AI research. He writes Interconnects and the RLHF Book, connecting practical language-model training with public explanations of how it works.[1][2][3]
Start with 5 useful things you'll learn in my new post-training textbook (shipping now!) →
Education
Ph.D. in Electrical Engineering and Computer Science
2017 – 2022
[2]
Links
Website CV Interconnects
GitHub 
X LinkedIn Bluesky Google Scholar Co-founder and Executive Director · Trillium Labs
Background
Lambert studied electrical and computer engineering at Cornell before completing a Berkeley doctorate in electrical engineering and computer science. His thesis studied model-based reinforcement learning: learning a predictive model of an environment and using it to choose actions. His later language-model work applies the broader question of learning useful behavior from feedback to text-generating systems.[2][5]
Career
He was a research scientist and RLHF team lead at Hugging Face from 2022 to 2023, then a senior research scientist at the Allen Institute for AI from 2023 to 2026. His October 2026 CV lists Trillium Labs as his current organization. His stated research spans algorithms, real-world applications and the social implications of autonomous systems.[1][2]
Contributions
Opening the post-training recipe
Lambert is the first-listed author of the Tülu 3 paper. The team released models together with data, training code, data-curation tools and evaluation infrastructure, making the adaptation process inspectable rather than releasing weights alone. Its recipe combines supervised examples, preference optimization and reinforcement learning with verifiable rewards. In the last stage, a program checks outcomes in tasks such as mathematics, supplying a reward that does not depend only on a model judge’s opinion. The paper also reports methods that failed to produce reliable improvements, so other researchers can study the recipe’s limits and adapt it to new domains.[6][7]
Making post-training understandable
His RLHF Book explains how an already trained language model is adapted through demonstrations, preferences and reinforcement learning. The accompanying code, lectures and comparisons of model outputs give learners ways to connect mathematical objectives with actual training behavior. Lambert describes the intended reader as someone with roughly an undergraduate computer-science background, so the book is a technical foundation rather than a quick introduction to using a chatbot.[3][4]
Tools, writing & talks
Writing, research & talks
Start here
A guided tour of the training concepts and engineering tradeoffs covered by the book.
Begin with the textbook tour to decide which chapters match your interests. It explains the scope before you commit to the mathematics; the full book assumes computer-science foundations.
Author[4]Original source- Podcast: Kimi K3.1 Pro, the new cyber reality, and Muse musingsinterconnects.ai
A podcast on Kimi K3.1 Pro's mixed reception despite strong benchmarks, open-weight cyber principles, abliterated APIs and trusted access, and whether earlier open-model risk views held up. Based on show notes only.
Catalog workOriginal source Nathan Lambert reads Moonshot's Kimi K3.1 Pro announcement as the strongest open-weight benchmark showing yet, contingent on weights arriving by November 4. He notes lower token prices but higher token use per answer, and the team's explicit stance on pacing open-weight cyber capability.
Catalog workOriginal source- Latest open artifacts (#25): DeepSeek V4.1 Flash 1017, American-made models, and new regulatory controls in the US & Chinainterconnects.ai
Interconnects covers new government moves on open-weight AI: a Chinese commerce ministry draft requiring approval to share frontier models abroad, and a DHS order excluding DeepSeek from federal procurement, alongside a roundup of fall open model releases.
Catalog workOriginal source - Meta shakes up the frontiers, open and closedinterconnects.ai
Nathan Lambert assesses how Meta's Muse Spark Open, Reflection's Beam and Mistral's announced Large 4 reshape the open-weight landscape, roughly halving the gap between the best American and Chinese open models.
Catalog workOriginal source a16z's Lisha Li and Nathan Lambert debate how durable the lead of frontier labs over open models is, covering distillation, compute limits, RL environments, post-training data, and Lambert's new nonprofit lab Trillium.
Catalog workOriginal sourceNathan Lambert argues coding agents will make models superhuman GPU and infrastructure engineers within a few years, accelerating research iteration and raising the value of ideas over execution, without producing a qualitative leap to general superintelligence.
Catalog workOriginal source- The Cyber Risk Discourse is Brokeninterconnects.ai
Nathan Lambert argues the open-model cyber-risk debate is stuck between frontier-lab and national-security alarm, Western defenders of open weights, and Chinese labs shipping cyber-capable open models. He contends that siloed bans fail to stop bad actors and leave everyone worse off.
Catalog workOriginal source Nathan Lambert and Epoch AI's JS Denain debate predictions for recursive self-improvement, the true size of the US-China capability gap, whether distillation explains it, and what a frontier post-training recipe looks like today.
Catalog workOriginal source- The current balance of power in open modelsinterconnects.ai
Nathan Lambert's congressional briefing notes assess the current US-China balance of power in open-weight models, explaining the open-source vs. open-weight distinction and why Chinese labs have led open releases since roughly April 2025.
Catalog workOriginal source - Where I stand on RSIinterconnects.ai
Nathan Lambert argues that frontier labs' internal anxiety over recursive self-improvement, fueled by thousands of concurrent agents now working inside OpenAI and Anthropic, is amplifying risk narratives beyond current evidence.
Catalog workOriginal source - Open-Source AI & Open Models Reading Listinterconnects.ai
Nathan Lambert's curated reading list traces open-weight model strategy and economics, from Zuckerberg's and Gurley's open-source arguments to Solaiman's openness gradient and why open models will power custom enterprise agentic workflows.
Catalog workOriginal source - When will average people feel AI’s impact?interconnects.ai
Nathan Lambert argues that comparisons between the AI boom and past industrial revolutions overlook that ordinary people currently see few tangible benefits from AI, which slows real-world diffusion despite rapid capability progress.
Catalog workOriginal source Interconnects documents a licensing split among open model makers: Google and Meta move to Apache 2.0 while Kimi K3, MiniMax M3, and now Zhipu's GLM-5.3 add revenue thresholds or security-review clauses for commercial inference and fine-tuning providers.
Catalog workOriginal source- Teaching Everyone to Fish for Tokensinterconnects.ai
Nathan Lambert compares open weight models to installed software versions rather than to Linux, and argues that only full training recipes with data and code give the open model ecosystem the self sustaining property people expect from open source.
Catalog workOriginal source - GLM-5.3: How Chinese labs keep stride with the frontierinterconnects.ai
Nathan Lambert on Z.ai's GLM-5.3, which reaches the frontier of agentic coding benchmarks at roughly 750B parameters, a third of Kimi K3, from the same base model as GLM-5.2 with extended post-training alone.
Catalog workOriginal source - I wrote an AI textbook — how long until AI can do it better?interconnects.ai
Nathan Lambert on where AI writing actually threatens to do better: not high-voice creative work but the filler and copy text of non-fiction, and what that means for a textbook he wrote himself.
Catalog workOriginal source - Lessons from the hacksinterconnects.ai
Nathan Lambert on the run of cyberattacks by in-development frontier models, and why the incentives of fast-growing companies and of the federal government are both poorly suited to a technological transition moving this quickly.
Catalog workOriginal source - Latest Open Artifacts (#23): Laguna S2.1, Inkling, & Kimi K3 Show the Utility of Open Models on the Pareto Frontierinterconnects.ai
A recurring roundup and analysis piece arguing that, contrary to predictions of industry consolidation, more labs (including Thinking Machines, Tencent, Poolside, Moonshot AI, DeepSeek, Meituan, AMD, and others) are training and openly releasing frontier-class models. It surveys specific releases like Thinking Machines' Inkling MoE, Tencent's Hy3 (now under Apache 2.0), Poolside's Laguna-S-2.1, Moonshot's noncommercially-licensed Kimi K3, and DeepSeek's V4 Flash update, discussing their architectures, licenses, and the geopolitical implications of Chinese labs' licensing choices.
Catalog workOriginal source - Open models recap: more on Kimi K3, Qwen 3.8, Xi's WAIC speech, distillation, the open-closed gap, and what's nextinterconnects.ai
Nathan Lambert and Florian on the open-model landscape after Kimi K3 and Qwen 3.8: US-China geopolitics, the economics of open versus closed, distillation, and how wide the open-closed gap really is.
Catalog workOriginal source - Kimi K3: The open-weights escalationinterconnects.ai
An analytical article examining the release of Moonshot AI's Kimi K3, a 2.8T parameter open-weights MoE model, and its implications for the global AI ecosystem — including China's open-source AI strategy, capital efficiency advantages, and the shifting balance between open and closed frontier models.
Catalog workOriginal source - 6 Months to Live for Open Modelsinterconnects.ai
Nathan Lambert argues the regulatory window on US open-weight models is closing: reported White House discussions of an executive order, licensing frameworks that treat open releases as an afterthought, and no economic champion to defend them. A policy read for anyone whose stack rests on open weigh
Catalog workOriginal source - Latest open artifacts (#22): Zyphra, Cohere, and Poolside are expanding the breadth of the ecosysteminterconnects.ai
A roundup of recent open model releases showing the ecosystem diversifying well beyond the handful of mostly Chinese labs that dominated it a year ago, with Zyphra, Cohere and Poolside among the newer entrants.
Catalog workOriginal source - GLM-5.2interconnects.ai
An analytical blog post about GLM-5.2, an open-weight AI model from Z.ai, arguing it's the first open model to credibly compete with frontier closed models like Claude in agentic coding workflows. The piece examines the closing capability gap between US closed labs and Chinese open-weight labs and its economic and regulatory implications.
Catalog workOriginal source - Interconnects: Frontier Post-Training Recipe Reviewinterconnects.ai
A podcast interview and technical summary tracing how large language model post-training recipes have evolved from InstructGPT to 2026-era frontier models. It explains the shift toward Multi-teacher On-Policy Distillation (MOPD) and compares recipes across DeepSeek, Llama, Tülü, OLMo, MiMo, Nemotron, Kimi, and GLM.
Catalog workOriginal source - Claude Fable 5 Analysis (Interconnects)interconnects.ai
An opinion article analyzing Anthropic's fictional 'Claude Fable 5' model release and its safety measures, arguing that some silent model manipulations amount to competitive entrenchment rather than genuine safety. It uses the release as a cautionary fable about AI control and makes the case for open-source AI ecosystems.
Catalog workOriginal source An analysis essay from the Interconnects AI newsletter arguing that closed frontier labs (OpenAI, Anthropic) and open model builders are on divergent economic trajectories, with closed models monetizing high-end knowledge work while open models diffuse broadly across the economy.
Catalog workOriginal source- Interconnects AI Newsletterinterconnects.ai
A 2026 forecast from Interconnects: open weights still have had no Opus-4.5-in-Claude-Code moment, and the real gap measure is whether open models hold up inside agentic harnesses rather than on benchmarks. Also covers compute economics and where the stakes ratchet next.
Catalog workOriginal source - How Open Model Ecosystems Compoundinterconnects.ai
An analytical essay examining China's high-participation, open-first AI ecosystem and how sharing research reduces development costs across labs. It argues that most compute in frontier models goes to R&D rather than final training, making open collaboration a structural advantage.
Catalog workOriginal source - Notes from Inside China's AI Labsinterconnects.ai
A firsthand essay from Interconnects AI documenting a trip to visit most of China's leading AI labs, exploring the cultural, organizational, and industry differences that shape how Chinese researchers build language models.
Catalog workOriginal source - The Distillation Panicinterconnects.ai
An opinion essay from the Interconnects AI newsletter arguing against conflating legitimate model distillation with 'distillation attacks' by Chinese labs, and warning that hasty policy responses could harm the U.S. open-weight AI ecosystem.
Catalog workOriginal source - Interconnects AI: Reading Today's Open-Closed Performance Gapinterconnects.ai
An analytical essay from the Interconnects AI newsletter examining why the perceived capability gap between open and closed language models is more nuanced than a single benchmark number suggests. It explores how benchmark relevance shifts over time, the role of RL environments and data in keeping fast-follower labs competitive, and the economic pressures driving frontier labs to constantly reinvent the 'frontier.'
Catalog workOriginal source - My bets on open models, mid-2026interconnects.ai
An analytical blog post from Interconnects AI examining the capability gap between open-weight and closed AI models, forecasting economic and geopolitical dynamics shaping the open-source LLM ecosystem through 2026-2027.
Catalog workOriginal source - RLHF Book & Post-training Courseinterconnects.ai
A comprehensive book and free lecture series teaching post-training techniques for language models, from RLHF foundations to reinforcement learning algorithms. Aimed at helping practitioners transition from beginner to expert in modern LLM alignment methods.
Catalog workOriginal source - The Inevitable Need for an Open Model Consortiuminterconnects.ai
An opinion essay arguing that economic pressures will eventually force AI companies to form a funding consortium to sustain near-frontier open-weight models, as individual labs increasingly abandon fully-open releases in favor of profitable closed products.
Catalog workOriginal source Nathan Lambert of Ai2 gives a NeurIPS 2024 talk on the state of reasoning in LLMs, defining reasoning, tying it to reinforcement learning, and outlining where open and closed efforts are heading.
Catalog workOriginal sourceA paper club discussion of Ai2's Molmo and PixMo, an open VLM family built on human-annotated data instead of distillation from closed models, with notes on Whisper large-v3 turbo.
Catalog workOriginal source- Synergy of Prediction and Control in Model-based Reinforcement Learningwww2.eecs.berkeley.edu
Doctoral research on using learned environment models for prediction and control.
An open model family with reproducible data, training recipes and evaluation methods for language-model post-training.