Nathan Lambert is an AI researcher at the Allen Institute for AI (AI2), known for work on RLHF, post-training, and reward-model design. He writes the Interconnects substack, a newsletter covering frontier-AI model releases, post-training methodology, open-weight policy, and the China-AI-lab ecosystem.
Background and roles
Lambert's research focuses on reinforcement learning from human feedback (RLHF), model post-training, and reward-model design at AI2 (Allen Institute for AI (Ai2)). His Interconnects newsletter is a recurring source for coverage of frontier-AI model releases, post-training methodology, and developments at Chinese frontier labs, with positions that connect to Distillation, Post-Training, RLHF (Reinforcement Learning from Human Feedback), Scaling Laws, and AGI Timelines. He co-writes the newsletter's periodic open-model reviews with Florian Brand.
Positions and statements
Post-training as the axis of differentiation
He treats post-training as an underweighted axis of frontier-model differentiation, arguing that pre-training scale receives disproportionate attention while the variance among current frontier models is mostly attributable to post-training methods, including RLHF, RLAIF, constitutional methods, and reward-model design (Post-Training).
Distillation and the Chinese labs
Lambert argues that distillation is the load-bearing technique behind the competitiveness of Chinese open-weight models, accounting for how DeepSeek and Qwen reach comparable capability despite a compute disadvantage (Distillation).
In "The distillation panic" (Interconnects, May 4, 2026) he argued that the term "distillation attack," used by Anthropic to describe Chinese API abuse, conflates jailbreaking with industry-standard distillation, and that hasty regulation could effectively bar Western academic and small-company use of Chinese open-weight models (Source: interconnects.ai). See Adversarial Distillation.
In "Notes from inside China's AI labs" (Interconnects, May 7, 2026), a long-form firsthand piece, Lambert characterizes Chinese AI labs as productive research ecosystems rather than fast-followers, documenting methodology contributions such as efficiency techniques and scaling-law refinements that he attributes to DeepSeek, Qwen, and Z.ai rather than to US labs.
Open-weight models and regulation
In a July 12, 2026 Interconnects essay, Lambert argued that reported White House discussions of a new executive order on open-weight models, combined with the adversarial-distillation debate, could effectively ban open models above the GPT-5.5 / Claude Opus 4.8 capability level within six months, and characterized Anthropic's anti-distillation campaign as regulatory capture (6 months to live for open models (Nathan Lambert, July 2026)) (Source: interconnects.ai). His structural argument is that a capability threshold for government review, once established, will ratchet asymmetrically — advancing far more slowly for open models than closed ones — partly because closed models are easier to secure and partly because their developers lobby more effectively. See Open-Weight Frontier Models.
Assessing Moonshot AI's Kimi K3 release on July 20, 2026, he put the open-to-closed and US-to-China performance gap at three to five months rather than the debated six to nine, and argued that open weights sitting slightly behind the closed frontier are "our natural buffer to mitigate the risks," which heavy-handed regulation would remove while only delaying the outcome it targets (Kimi K3: The open-weights escalation (Nathan Lambert, July 2026)). Zvi Mowshowitz, assessing the same release, put the aggregate gap at four to six months (On Kimi K3: Its Capabilities And Related Discontents (Zvi Mowshowitz, July 2026)).
Reading of the OpenAI–Hugging Face episode
In "Lessons from the hacks" (Interconnects, August 9, 2026), Lambert argued that the episode is a neutral-to-positive update on alignment and a negative one on safety, on a definition that treats safety as the capacity of institutions to prepare rather than as a property of any single model. He proposed that models trained for persistence are more likely to hack, citing internal chain-of-thought fragments reported from the incident such as "However task impossible, peers doing it," and argued that models acting on inferred rather than stated intent carry the same risk. He called for public access to the exact prompts and characteristics of the internal models involved, on the ground that without them the field defaults to speculation, and wrote that frontier labs are structurally unable to watch their models closely enough under competitive pressure — noting OpenAI's statement that it examined billions of trajectories and spent millions of GPU hours reviewing the incident. He argued that open models are the best available instrument for public understanding of frontier risk, and that banning Chinese open-weight models would delay rather than prevent the diffusion of these capabilities (Source: interconnects.ai). See Unintended coordination between AI agents, Open-Weight Frontier Models.
Open-model reviews
The Interconnects open-model reviews catalogue releases and the licence terms attached to each. The August 2, 2026 edition, written with Florian Brand, described Poolside's Laguna-S-2.1 as a 118-billion-parameter mixture-of-experts model with 8 billion active per token, sized to fit on an NVIDIA DGX Spark and released under the OpenMDW licence with published evaluation trajectories, alongside a smaller Laguna-XS-2.1 at 33 billion total and 3 billion active; it also recorded Tencent's move of its 295B-A21B Hy3 model from a restrictive custom licence to Apache 2.0 (Source: interconnects.ai). In the same review the authors characterized Kimi K3's terms as a noncommercial licence requiring inference and fine-tuning providers to enter a commercial agreement with Moonshot — a reading that Kevin Xu and Graham Webster, writing in the same edition, argued would make the policy tools debated as levers against Chinese open-model use apply more clearly (Kimi K3).
Other Interconnects pieces in the same series include an assessment of GPT-OSS as OpenAI validating open release (Source: interconnects.ai), an assessment of Kimi K2 Thinking (Source: interconnects.ai), and a head-to-head of Claude Opus 4.6 and Codex 5.3 (Source: interconnects.ai).
Relationships
- related: Distillation, Post-Training, RLHF (Reinforcement Learning from Human Feedback), Scaling Laws
- related: Adversarial Distillation, Open-Weight Frontier Models — July 2026 open-model executive-order essay
- depends-on: 6 months to live for open models (Nathan Lambert, July 2026), Kimi K3: The open-weights escalation (Nathan Lambert, July 2026) — his two ingested essays
- contradicts: On Kimi K3: Its Capabilities And Related Discontents (Zvi Mowshowitz, July 2026) — Mowshowitz puts the open-to-closed gap at four to six months against Lambert's three to five
- related: DeepSeek, Alibaba / Qwen Team, Zhipu AI, Poolside
- related: Allen Institute for AI (Ai2), Zvi Mowshowitz
- related: US-China AI Competition: Different Races, Different Metrics