Interconnects (interconnects.ai)

Interconnects is the newsletter of Nathan Lambert, a senior research scientist and post-training lead at the Allen Institute for AI (Ai2), where he trains the fully open olmo models. It goes out one to three times a week on Substack: essays, reviews of new models, interviews with researchers, and surveys of open model releases. Some essays and the monthly "artifacts" roundups of open models and datasets are for paid subscribers only. Most posts also go out as audio in a podcast feed.

What sets it apart from other AI commentary is that the author trains models for a living. His research history shows in the topics: RLHF and reward modelling (Tülu, RewardBench, Zephyr-Beta, the Open LLM Leaderboard), the first textbook on RLHF (rlhf-book-synthetic-data-distillation is one chapter of it), and since 2025 reasoning models. The most-read posts are close readings of frontier releases — the DeepSeek R1 recipe, o3's over-optimization. He also gives talks on the same material; lambert-china-ai-ecosystem-open-model-gap is his 2025 account of how Chinese open models overtook Llama.

Since 2025 the blog has also been an advocacy platform. Lambert runs the ATOM Project (atom-project-american-truly-open-models), a campaign for American fully open models, and a growing share of posts are policy-facing: whether open models should be banned, how Chinese labs keep pace, and the distillation fight. Keep that in mind when reading it as evidence. He is a participant in the argument, and he is candid about revising himself — open-source-ai-reading-list records him walking back his 2025 confidence that DeepSeek did not distil o1.

Not to be confused with Kevin Xu's Interconnected newsletter, which the reading list also cites (chinese-open-source-a-definitive-history).

Ingested articles

• open-source-ai-reading-list — 2026-09-11 — Annotated reading list on open models: strategy, China, the gap, cyber risk, distillation
• what-comes-next-with-open-models — 2026-03 — Open models as the complement to closed ones, the base for enterprise agent workflows
• open-models-in-perpetual-catch-up — 2026-02 — Why open models will keep trailing closed models in performance
• open-and-closed-models-are-on-different-exponentials — 2026-06 — Open and closed models follow different adoption curves
• kimi-k3-open-weights-escalation — 2026-07 — Kimi K3 and what a Chinese open frontier model does to the argument
• glm-5-2-step-change-for-open-agents — 2026-06 — GLM-5.2 as the step change for open agentic models
• why-i-build-open-language-models — 2024-10 — Lambert's case for building open models as a research practice
• banning-open-source-ai-would-be-a-mistake — 2026-06 — With Kevin Xu: education, innovation and competition as the case against a ban
• six-months-to-live-for-open-models — 2026-07 — "Vibe regulation" and the coming clash over frontier open weights
• notes-from-inside-chinas-ai-labs — 2026-05 — How Chinese labs describe building models, and how their industry differs
• glm-5-3-how-chinese-labs-keep-stride — 2026-08 — How Chinese labs keep pace with the frontier
• how-much-does-distillation-matter-for-chinese-llms — 2026-02 — Distillation helps Chinese labs without explaining away their work
• the-distillation-panic — 2026-05 — The political panic over distillation is not grounded in evidence
• how-distillation-is-used-today — 2026-07 — Distillation's current uses and uplift for open models (natolambert.substack.com)
• frontiers-in-synthetic-data — 2024 — Synthetic data and SFT distillation as the dominant post-training tools