What Interconnects is
Interconnects is Nathan Lambert's newsletter and commentary platform covering the AI/ML landscape, with a particular focus on open-weight models, post-training techniques, and the structural competition between open and closed frontier systems. It operates explicitly as a tier-2 source: a synthesis and analysis layer that reads primary lab announcements, research papers, and practitioner experience, then distills them into opinionated, technically grounded takes for a practitioner audience.
Lambert's background — which the events bundle places at the intersection of RLHF research and open-model advocacy — gives the platform a distinctive vantage point. Interconnects is not neutral aggregation; it has recurring theses it tests against new evidence, and it is willing to push back against dominant narratives from frontier labs, policy circles, and the AI safety community alike.
Why it matters
For practitioners tracking the open-weights ecosystem, Interconnects functions as a high signal-to-noise filter. The recurring "Latest open artifacts" roundup series (issues #19, #20, #21, and #22 appear in this bundle) provides structured coverage of a release cadence that has become genuinely difficult to follow — issue #21 alone covers Gemma 4, DeepSeek V4, Kimi K2.6, MiMo 2.5, and GLM-5.1, characterizing the period as a "flagship-after-flagship" moment. Issue #22 extends coverage to second-tier labs — Zyphra, Cohere, and Poolside — examining the motivations behind their open releases and what it means for ecosystem breadth.
Beyond aggregation, the platform's analytical framing shapes how practitioners think about model selection and deployment strategy. The argument that open and closed models are "on different exponentials" — following distinct capability and value curves rather than competing head-to-head — is a durable framing that has implications for how enterprises and researchers should approach model choice.
Recurring analytical threads
Several theses recur across the bundle's events:
Open vs. closed dynamics. Multiple pieces examine the performance gap between open-weight and closed frontier models, what drives it, and under what conditions open models can close or circumvent it. "Open Models in Perpetual Catch-Up" frames distillation as a structural catch-up mechanism while identifying conditions where open models can win outright. "Reading today's open-closed performance gap" interrogates the benchmark variables that produce the single numbers dominating public discourse.
Distillation discourse. "The Distillation Panic" critiques the term "distillation attacks" as misleading or alarmist for what is simply training on frontier model outputs. A follow-up piece, "How much does distillation really matter for Chinese LLMs?", reacts to Anthropic's framing and interrogates how much capability transfer actually explains Chinese LLM progress — a more empirical take on the same debate.
Chinese open-weights ecosystem. Interconnects covers Chinese labs with unusual depth. Issue #19 tracks Qwen 3.5, GLM 5, and MiniMax 2.5 as a frontier push; "How Open Model Ecosystems Compound" argues that China's open-first, high-participation ecosystem creates structural compounding advantages; and a firsthand account from visits to leading Chinese AI labs offers rare insider perspective on research culture and strategic direction.
Post-training and evaluation. Interview #18 with Finbarr Timbers covers frontier post-training recipes — RLHF, preference optimization, and related techniques. "Opus 4.6, Codex 5.3, and the post-benchmark era" argues that traditional benchmarks are no longer sufficient for distinguishing frontier model capabilities, reflecting a broader industry shift toward task-specific evaluation.
AI safety and policy. Interconnects engages critically with safety narratives. A piece on "Claude Fable 5 and AI safety power politics" examines how frontier labs construct and communicate safety claims. "Claude Mythos and misguided open-weight fearmongering" pushes back on narratives framing open-weight releases as uniquely dangerous. "Welcome to the AGI era of AI governance" argues that AI governance has crossed a one-way threshold it was unprepared for. An op-ed co-authored with Kevin Xu argues directly against banning open-source AI.
Self-improvement and AI risk. "Lossy self-improvement" takes a nuanced middle-ground position: AI self-improvement is real, but inherent lossiness prevents the discontinuous fast-takeoff scenarios common in AI safety discourse.
Format and cadence
The platform mixes several formats: numbered recurring roundups ("Latest open artifacts"), numbered interviews, long-form analytical essays, forward-looking monthly commentary ("Some ideas for what comes next, May 2026"; "My bets on open models, mid-2026"), and occasional op-eds for broader audiences. This variety lets it serve both as a tracking resource for specific releases and as a venue for durable analytical arguments.
Recent developments
The most recent events in the bundle show Interconnects tracking the frontier in real time: GLM-5.2 is flagged as a step-change capability threshold for open-weights agentic models; issue #22 surveys Zyphra, Cohere, and Poolside's ecosystem contributions; and the platform continues to engage with policy debates around open-source AI regulation. The "AGI era of AI governance" framing, published in mid-June 2026, suggests Lambert views the current moment as a structural inflection point — not just a capability milestone — with governance implications that the field has not yet caught up to.
Where it's heading
The bundle's arc suggests Interconnects will continue to serve as a practitioner-level counterweight to both frontier-lab PR and AI safety alarmism, with deepening coverage of the Chinese open-weights ecosystem, post-training technique evolution, and the governance questions that accompany rapid capability gains. Its value is precisely that it is opinionated, grounded in technical fluency, and willing to name what it thinks is wrong with dominant narratives.
