What I’ve been building: ATOM Report, post-training course, finishing my book, and ongoing research
Interconnects Nathan Lambert
Nathan Lambert dropped four projects at once: an open-model tracker, a finished book, a free course, two papers. His new metric shows China's open models catching up fast.
Based on reporting by Interconnects, Nathan Lambert — read the original for the full story.
Summary, retelling and take written by AI under human oversight; images are AI-generated illustrations. How we work · Report an error
Nathan Lambert has apparently decided that writing one blog post at a time is inefficient. His latest roundup bundles a research report, a finished textbook, a new course, and two academic papers, all shipped in the same stretch of weeks. Taken together, they read less like a grab bag and more like someone quietly building the infrastructure for how the industry understands open-weight AI.
The centerpiece is the updated ATOM Report, the technical backbone behind The ATOM Project's push for more U.S. investment in open models. Lambert and collaborator Florian refreshed their Relative Adoption Metric, or RAM, a size-normalized way to track how fast a model gets downloaded relative to its historical peers. A RAM score above 1 means a model is pacing to land among the ten most-downloaded releases ever in its weight class. Run through recent Chinese releases from Moonshot, Z.ai, and MiniMax, the metric shows a mid-tier of Chinese labs quietly punching above their weight, while GPT-OSS and the freshly launched Gemma 4 show unusually strong early adoption on the U.S. side.
On a completely different track, Lambert's RLHF Book finally shipped to production with Manning after nearly two years of work — he bought the rlhfbook.com domain back on May 20, 2024. Print copies are roughly two months out, and it's up for preorder now, cheaper direct through Manning than on Amazon. He calls it the book he wished existed when he started in post-training, which is a fair description of a field that until recently ran mostly on scattered blog posts and Discord threads.
Rather than stop at a book, he's turning it into something closer to a curriculum. A free lecture series is rolling out on YouTube, four videos in so far, covering everything from RLHF fundamentals to implementing policy-gradient RL for language models, with community Q&A sessions slotted between lectures. It's a bet that as models get harder to reason about, teaching people to actually read research matters as much as writing it.
The research itself hasn't stopped either. Two new papers, one on the gap between single-turn and multi-turn model capabilities and another framing reinforcement learning as a meta-learning problem where models learn from their own past attempts, point toward where Lambert says his attention has drifted: agents, and the tricky interface question of what information to surface to a user mid-task. He's now headed to China and possibly Washington D.C. to see how each side is actually reading the state of the race, which feels like the right next move for someone who just spent months trying to measure it from the outside.
My take — AI-written commentary, not fact-checked reporting
I like that someone is actually trying to put numbers on the open-model race instead of just vibes and screenshots of leaderboards — RAM is a genuinely useful idea. But the bigger tell here is that a researcher felt the need to build a tracker, a book, and a course just to keep the field legible; that's a sign the ecosystem is moving faster than any single company's PR team can honestly explain, and open models badly need people doing this unglamorous accounting work.
Read more about this at: Interconnects