The Man Making Sure AI's Future Stays Open

Nathan Lambert showed up to UC Berkeley's PhD program in 2017 with a resume that made zero sense for AI: MEMS devices, high-energy physics lasers, a stint doing battery engineering at Tesla. He wanted to work with Sergey Levine or Pieter Abbeel - the AI royalty of Berkeley - and both said no. So he did what any sensible person would do: he taught undergraduate courses to pay his own way through a PhD, built his own foundations in machine learning, and quietly turned himself into one of the most cited researchers in reinforcement learning from human feedback.

The backstory matters because Lambert's outsider arc is not incidental to his work - it is his work. He spent years solving the actual hard problem: how do you take a language model trained to predict the next word and teach it to behave in ways that humans actually want? The field calls this post-training. The rest of the internet calls it "why does ChatGPT not tell me how to make poison." Lambert built the infrastructure, wrote the textbook, and explained it to anyone willing to read 3,000 words over breakfast.

Right now he is Senior Research Scientist and Post-Training Lead at the Allen Institute for AI (Ai2), the non-profit research organization funded by the Paul Allen estate. His job is to make sure OLMo - Ai2's flagship open-source language model series - is not just capable, but genuinely open. That means releasing model weights, yes. It also means releasing training data, intermediate checkpoints, training logs, technical decisions, and the reasoning behind every choice made along the way. In a field where even "open" is often a marketing term, Lambert's team actually publishes the receipts.

The RLHF path started at Hugging Face, where Lambert joined as Lead Scientist heading up the RLHF team. He was recruited by Douwe Kiela - which, in AI terms, is roughly equivalent to being tapped by a scout who has an eye for talent. At Hugging Face he co-built TRL (Transformer Reinforcement Learning), an open-source library that became the standard toolkit for training language models with human feedback. He shipped Zephyr-Beta, an RLHF-trained model that punched well above its weight class. Then he went to Ai2 and did it again, at bigger scale, with fuller transparency.

Tulu 2, co-authored under his watch, was the first publicly documented instance of DPO - Direct Preference Optimization - scaling to 70 billion parameters. Tulu 3 pushed further. OLMo 2 came in 7B, 13B, and 32B variants, all fully open. OLMo 3's 32B model was described at release as the best open-source language model available. These are not press releases. These are reproducible results with training code, data, and logs attached.

The Newsletter That Refused to Stay Simple

Before any of the Ai2 work, before TRL, before Zephyr - there was Interconnects AI. Lambert started writing it as a hobby, a place to think in public before ChatGPT turned AI into the only dinner conversation anyone wanted to have. When ChatGPT launched in late 2022, he made a decision: stop treating it as a side project. The newsletter became a weekly commitment. Then more than weekly. Now he publishes on an almost feverish schedule - essays, model breakdowns, interviews with researchers, surveys of the open model ecosystem.

Interconnects AI earned over 1.2 million page-views in 2024 and crossed 3.5 million in 2025. It is read by engineers at frontier labs, investors making multi-million-dollar bets, and policy people trying to figure out what any of this actually means. The appeal is specific: Lambert writes like someone who has run the experiments, made the mistakes, and is now patiently explaining what actually happened rather than what the press release says happened. More than 300 posts, almost entirely by one person, with no venture funding and no marketing team.

The newsletter has a dedicated interview series - conversations with researchers tracking the technical shifts that often don't make it into mainstream coverage. Lambert has appeared on Lex Fridman's podcast, Latent Space, The Cognitive Revolution, ChinaTalk, and Understanding AI with Tim Lee. He doesn't perform expertise. He just has it.

The Book He Had to Write

There was no good textbook on RLHF. Lambert decided to write it. The RLHF Book, published through Manning Publications, is in Early Access as of 2026. It covers the full pipeline - instruction tuning, reward modeling, rejection sampling, RL optimization, and direct alignment algorithms like DPO. An ArXiv companion paper was published in April 2026 (arxiv.org/abs/2504.12501). There is also an accompanying free course with lectures. The book draws on philosophy, economics, and optimal control alongside the machine learning - which reflects Lambert's own path through multiple fields before he landed in AI.

An accompanying GitHub repository at github.com/natolambert/rlhf-book has the supporting code open for anyone to use, fork, and build on. Because of course it does.

The Thing About Open Source

The word "open" in AI is doing a lot of heavy lifting right now. Many companies use it to mean "we released the weights and hope you don't look too closely at the rest." Lambert's work at Ai2 means something different: full openness, where every training decision, every data source, every intermediate checkpoint is documented and shared. The Open Instruct codebase at github.com/allenai/open-instruct is a collaborative, public post-training stack anyone can run, inspect, or improve.

Lambert is not naive about the tradeoffs. He writes clearly about where open models fall short, where closed systems have real safety advantages, and where the frontier labs are running experiments nobody else can replicate. What he argues for - consistently, over hundreds of thousands of words - is that the baseline for "open" should be higher, that researchers should be able to actually reproduce results, and that releasing weights without releasing data and training code is a form of theater.

That's a position that takes courage to hold when the funding flows toward closed systems and the incentives push toward locking up competitive advantages. Lambert holds it anyway, and publishes the evidence on Substack twice a week.