Interconnects, the Substack of former AI2 post-training lead Nathan Lambert, is the inside-the-labs newsletter for the models and reasoning layer under agentic development, minus the hype.
He explains the research and the open-model ecosystem from someone who built it, which makes him the direct bridge between the frontier labs and the engineer choosing a model or an agent today.
What it is #
A Substack newsletter, website, and interview podcast by Nathan Lambert, a machine learning researcher who was a post-training lead at the Allen Institute for AI (Ai2) and co-led the OLMo open-model effort.
The newsletter describes itself as the cutting edge of AI from inside the frontier labs, minus the hype.
His Get Good at Agents (2026-01-21) piece and his
Claude Code reviews document how he actually operates coding agents in his own RL research.
Status #
Active and prolific as of 2026-10-10.
The newsletter reports over 85,000 subscribers as of 2026-10-04, and his most recent essay is “I expect rapid progress but not towards general superintelligence” (2026-10-09, his case that the visible acceleration is in engineering and infrastructure capability, that coding agents make research easier and good ideas more valuable than execution, and that pretraining research could be automated in two to three years without models changing in nature), which follows “The Cyber Risk Discourse is Broken” (2026-10-06), his argument that the open-weights cyber-risk debate is heading toward lose-lose policy, that closed-model APIs cause most documented attacks to date, and that banning open weights without also banning closed APIs would widen the offense-defense gap; it follows “The current balance of power in open models” (2026-09-21), the expanded form of testimony he prepared for Congress, and a podcast with JS Denain of Epoch AI (2026-09-22), after the RSI-skeptic essay “Why I still haven’t bought into true RSI” (2026-09-19).
He released the
RLHF Book (‘26) with a companion post-training course, and his site links to open-model tools he maintains: the Artifacts Hub, an Adoption Dashboard, and the ATOM Project.
The “currently doing something new” tease after leaving Ai2 has landed: on 2026-10-01 he co-founded
Trillium Labs with Tom Zick, a nonprofit for the open science of frontier AI that will build fully open post-training recipes and open infrastructure to study RSI, reward hacking, and multi-agent systems, with early support from Halcyon Futures and Schmidt Sciences.
Strengths #
- He writes about post-training, reasoning, and open models from direct authorship rather than press coverage, so the technical claims are reliable.
- He uses the coding agents under discussion in his own research work, documenting real multi-agent and RL workflows rather than demos.
- The focus on open models and candid comparisons gives readers a usable lens for model selection without relying on benchmark marketing.
- The RLHF Book and course are a rare, grounded curriculum for the post-training layer that most practitioners never get.
Cautions #
- The newsletter is opinion-driven and fast-moving, so takes can shift and need re-checking against primary sources.
- Some posts are analysis and commentary rather than new findings, and readers must separate signal from churn.
- The writing is research-oriented and assumes familiarity with fine-tuning and RL concepts.
- He is an open-model advocate, so the open-versus-proprietary framing is a side he argues from, even when balanced.
Pricing #
Free to read, reader-supported via paid subscriptions. The RLHF Book is paid; parts of the course and companion materials are free.
Compared to #
- Lilian Weng: both explain the model and reasoning layer, but Weng surveys the broader research canon while Lambert tracks the current open-model and post-training state.
- Chip Huyen: the model-training microview versus the application-systems macroview; Lambert is upstream of where Huyen’s survey starts.
- Andrej Karpathy: both are researcher-writers who set vocabulary, but Lambert is high-cadence and tool-engaged while Karpathy is low-frequency and conceptual.
Bottom line #
Recommended for the engineer who wants to understand the reasoning-and-post-training layer behind the agents they use, and to pick open models with good information. Not for anyone who wants a mostly-vendor-neutral or application-level overview; this is the model-side view, argued from an open-model vantage.
Top 5 recommended reading #
Get Good at Agents - His documented personal workflow for scoping and running coding agents, the piece most directly useful to engineers who use them daily.
Reinforcement Learning from Human Feedback - His RLHF and post-training textbook with its companion course, the durable curriculum behind everything he writes about the model layer.
Changes #
- 2026-08-29 - Created as the model-and-post-training band of the people and publications category expansion.
- 2026-09-16 - Subscriber count moved to over 83,000 (from over 82,000), with no new posts since the September 8 to 11 cluster.
- 2026-09-20 - New essay “Why I still haven’t bought into true RSI” (2026-09-19) ended the quiet stretch after the September 8 to 11 posts.
- 2026-09-22 - Two posts after the RSI essay: “The current balance of power in open models” (2026-09-21, expanded Congressional testimony) and the Epoch AI podcast with JS Denain (2026-09-22).
- 2026-09-24 - Added the Top 5 recommended reading section.
- 2026-09-27 - Subscriber count moved to over 84,000; no new posts since the September 21 to 22 pair.
- 2026-10-04 - Window audit: added the Trillium Labs launch (2026-10-01), the nonprofit he co-founded with Tom Zick for open post-training science, resolving the “doing something new” teaser; subscriber count refreshed to over 85,000 as of 2026-10-04.
- 2026-10-07 - New essay “The Cyber Risk Discourse is Broken” (2026-10-06) now leads the newsletter, his case that the open-weights cyber-risk debate ignores trade-offs and that closed-model APIs are the documented attack vector to date; added to References (URL fetched this run).
- 2026-10-10 - New essay “I expect rapid progress but not towards general superintelligence” (2026-10-09) now leads the newsletter, his engineering-acceleration-not-RSI argument with the two-to-three-year pretraining-automation prediction; added to Status and References (URL fetched this run).
See also #
- Model Selection for Coding Tasks - the decision his model-and-reasoning coverage feeds
- Agentic Coding Tools Landscape - the harnesses he uses and evaluates in his own research
- Lilian Weng - the complementary research-survey voice
- Scaling the LLM Agent Company - the supervision and model questions his multi-agent RL work raises
References #
https://natolambert.com/ - his site with background, the RLHF Book, and the open-model tools
https://aiweekly.co/alerts/nathan-lamberts-rlhf-course-reaches-tool-use-and-agents - a third-party note that grounds his course contents and its stated limits
https://blog.trilliumlabs.org/cp/218529442 - the Trillium Labs launch announcement (2026-10-01), grounding the nonprofit, its co-founder Tom Zick, and its open post-training mission