Nathan Lambert publishes definitive open-models reading list: gap now just 4-6 months
Interconnects (Nathan Lambert) · rss · 2026-09-11
Nathan Lambert (Interconnects) published a continuously updated reading list covering the best writing on open models from recent years, organized into several sections:
- Foundation: Why open models exist and their risks — Bill Gurley on open-source strategy, Zuckerberg on Llama 3, Irene Solaiman's "gradient of release" framing, Thinking Machines Lab's A Safe Path to Open Weights, and early marginal-risk research.
- US-China competition: The ATOM Project, Kevin Xu's history of Chinese open source, Lambert's notes from inside China's AI labs, plus regulatory scrutiny of DoorDash, Airbnb, Cursor, and Apple over Chinese model usage, and Perplexity's DeepSeek R1 adoption and Thomson Reuters moving from Claude to Qwen.
- Technical details: The open-closed gap has narrowed to roughly 4-6 months, with leading open models all from Chinese labs since 2024, per SemiAnalysis evaluations and Z.ai's "open-source within hours" playbook. Includes a deep dive on the 2026 distillation debate and the recent paper on stealing reasoning traces from proprietary LLM APIs.
- Cyber & risks: Joshua Saxe's argument that bad actors can't be reliably cut off from open models and that the US needs a coherent national AI cybersecurity policy.
It's the most complete one-stop entry point for understanding the open-model landscape.
More from AGI Musings
- AI filmmaker argues hybrid production keeps gatekeepers in charge, backs 100% AI filmmaking — taherdhanera · 2026-09-11
- Drexler on the HuggingFace incident: system structure, not model alignment, drove the change — mattbeane · 2026-09-11
- CNBC examines risks of recursive self-improvement, quoting Conitzer — conitzer · 2026-09-11
- Gary Marcus: strategy is community pressure to stop AI labs getting richer and more powerful — GaryMarcus · 2026-09-11
- 'Most domains are shallow' sparks pushback: many fields are surprisingly deep for AI — binarybits · 2026-09-11
- AI safety insiders warn: social incentives push researchers toward exaggerated doom estimates — natolambert · 2026-09-11