Nathan Lambert is an AI researcher specializing in reinforcement learning from human feedback (RLHF) and post-training of language models. He spent several years at the Allen Institute for AI (Ai2) as a senior research scientist and post-training lead, where he was a core contributor to the OLMo family of fully open models as well as Tülu and RewardBench. He writes the widely read Interconnects newsletter and authored an online textbook on RLHF. He departed Ai2 in mid-2026 and, as of this writing, has not announced his next role, saying he plans to keep working on fully open post-training and medium-sized open models.
Recent
-
2026-07-27 Kimi K3 weights released at midnight UTC today; community benchmarking on the actual weights begins; his pre-release estimate of “3-5 months” open-to-closed frontier gap – and the 51% hallucination rate flagged by Artificial Analysis – now both face independent replication. interconnects.ai
-
2026-07-24 Published “Kimi K3: The open-weights escalation” on Interconnects: argues the open-to-closed frontier gap has moved from “6-9 months” to “3-5 months”; calls K3 (#3 on Artificial Analysis Intelligence Index, #1 in Frontend Code Arena) the strongest open model yet; frames the accelerating Chinese open-weight release cadence as a structural shift, not an anomaly. interconnects.ai
-
2026-07-05 Posted “GLM-5.2 is the step change for open agents” on Interconnects — argues GLM-5.2 is the “DeepSeek moment” for open agents: first open model competing with frontier proprietary on agentic coding, with the open-to-closed capability lag now ~6-8 months. post · tweet
-
2026-06-22 Released TMax (arXiv 2606.23321) with Ai2: open RL recipe for terminal agents shipping 9B and 27B weights, training data, and reproduction code. TMax-27B scores 42.7% on Terminal-Bench 2.0, best among public models. Key method: combining difficulty control, personas, and verifier diversification to cheaply generate terminal environments for RL. arXiv · Substack · X
-
2026-06-21 Noted that Chinese labs are “hitting high peaks with much less compute” and shouldn’t be discounted, while still maintaining US frontier models are overall ahead. X
-
2026-06-15 Confirmed as Global CTO of AI at the Linux Foundation, focusing on open model governance and fully open post-training work. Interconnects
-
2026-06-13 Posted Substack reflections on “the current state of the Claude Fable saga,” noting a cluster of overlapping problems operating on different timelines — some fixable, some without easy answers. Substack note
-
2026-06-12 Anthropic issued a formal apology to WIRED citing “the wrong tradeoff” — directly referencing the researcher backlash he led. The company confirmed the covert steering-vector degradation is now removed; requests that would have been silently throttled instead visibly fall back to Opus 4.8. Substack
-
2026-06-10 Published a detailed Interconnects post on Fable 5’s safety classifier design: “An AI model that gets less intelligent automatically without notifying me is categorically misaligned AI.” Argued that covert routing blocks legitimate open researchers while bad actors work around it. Anthropic reversed the hidden researcher throttle the same day. (Interconnects · X)
-
2026-06-09 Called out Anthropic publicly: “I don’t really want to have to go to bat against Anthropic, but they’ve just been unnecessarily antagonistic to all of China, then not so subtly to open weight models, and now more broadly open AI research. What’s next on the list?” Posted on the day Fable 5 launched. (tweet)
-
2026-06-06 Building out a deeper understanding of Chinese AI labs. Said he’ll share more soon. (tweet)
-
2026-06-06 Flagged rapid political pressure to ban AI “distillation attacks,” warning that early legislation could trigger serious unintended second-order effects. Substack