Interconnects publishes a reading list on open models and the US-China AI race
An Interconnects post presents a curated reading list on open-weight AI models. Its first-person author, Nathan Lambert, says the list was put together while prepping for public-audience and policy-facing writing on open models, and calls it their own pick of the best writing on the subject from the last few years, meant to bring a newcomer fully up to speed on the state of the field. The author invites readers to suggest more pieces in the comments and says the list, last updated 13 September 2026, will keep growing over time. The material is organized under three headings: Foundation, US-China Competition, and Technical Details.
Foundation covers why companies release open models and how seriously they take the safety trade-offs. It points to Bill Gurley's history of open-source software strategy applied to AI, Mark Zuckerberg's stated rationale for releasing Llama 3 as open, and Irene Solaiman's argument for treating openness as a gradient of license terms, running cost and data access rather than a binary choice. Several entries are the author's own past pieces, arguing that open models will complement rather than replace strong closed models by powering custom enterprise agentic workflows, that open models are stuck in perpetual catch-up on raw performance, and that adoption of open and closed models is following different curves. Christian Catalini's piece works through the economics of who captures value between open and closed AI, drawing on the history of intellectual property and the distillation debate. On safety, the list cites Thinking Machines Lab's case for releasing powerful open weights responsibly, an earlier Kapoor and Bommasani paper finding that text-focused LLMs only marginally increased documented potential risk, and Florian Brand's argument that closed models' own safety guardrails are already bypassed often enough to cause real harm before the hypothetical risks of open weights show up. Shayne Longpre's paper on the rapid decline of the open AI data commons gets a mention too, alongside the rise of strong Chinese open releases like Kimi K3 and GLM-5.2 and a Golden Gate Institute for AI talk by Nathan Lambert recapping 2025's story of the open-model gap. An optional closing block points to three resources for tracking adoption: the ATOM Report, an Interconnects adoption dashboard, and an 'Artifacts Hub' listing the ecosystem's most important models.
US-China Competition sets out why the author thinks America needs to keep investing in open models given China's rise, arguing open models foster education, innovation and competition, values the author frames as core to the American approach. One entry, titled 'six months to live for open models,' warns that vague, reactive 'vibe regulation' could soon produce a clash with, or an outright ban of, frontier open models. Optional links to fully open technical reports, Pythia and three generations of Olmo, are offered as reference points for what full model openness looks like in practice. On the Chinese side, the list draws on Kevin Xu's history of Chinese open source and his case for China's structural advantages in the space, plus a piece on how Chinese labs operate from the inside and another arguing that American open-weight labs themselves struggle to match Chinese labs on fair benchmark comparisons, one example the author gives of how well Chinese labs keep pace with the frontier. The section also covers regulatory fallout from the trend: citing CNBC, Bloomberg, Semafor and Reuters reporting, it says lawmakers have probed DoorDash, Airbnb, Anysphere/Cursor and Apple over their use of Chinese models, without saying what came of those inquiries. In the other direction, it points to Western companies that have publicly switched to Chinese open models to cut costs, citing Perplexity's adoption of DeepSeek R1 and Thomson Reuters building on Qwen to move off Claude.
Technical Details opens with the list's headline technical claim: the open-closed performance gap has narrowed in recent years to roughly four to six months, and the leading open models have all come from Chinese labs since about 2024. It backs that with a SemiAnalysis evaluation concluding open models have been closing on the closed frontier over time, further gap-tracking data from Epoch AI and Artificial Analysis, and an independent analysis by Håvard Tveit Ihle. It adds that open models sit on the Pareto cost frontier, the best performance available at a given price, without reaching the absolute performance frontier, pointing to DeepSeek V4 Flash as an example. On how fast Chinese labs move, it quotes a 2025 remark from the product lead of Z.ai: 'Get it out fast. We open source it within a few hours.' On cyber risk, it cites two essays arguing that banning open models would not remove bad actors' access to the underlying capability and that governments need a coherent, proactive national AI cybersecurity policy instead, alongside a separate essay arguing that trying to control access to AI above some capability threshold is the wrong way to manage misuse altogether.
The list calls distillation, training a model on another model's output tokens, 'the single most eventful debate around open models in 2026,' and gives it the most sustained treatment of the three sections. It points to a 2026 post-training textbook chapter for background on synthetic data, then to the author's own argument that distillation helps Chinese labs without taking away from their underlying innovation. It cites a September 2026 Anthropic report, 'Detecting and countering misuse of AI,' which it says documents at-scale use of Anthropic's products by banned parties, combining distillation via SFT data with extensive, undisclosed routing of Claude into their own products and services. It also cites a paper by Panfilov, Schmotz, Shumailov and co-authors describing techniques for systematically extracting reasoning traces, which the list calls the crucial part of modern training, from proprietary LLM APIs; Anthropic has confirmed Chinese labs used the technique, though the list does not say which ones. Against that backdrop, the author argues that the broader political panic treating distillation as the sole reason Chinese models sit near the frontier is not supported by the evidence. The list closes the thread by revisiting the 2025 dispute over whether DeepSeek R1 was distilled from OpenAI's o1: it says there is no clear evidence for that, and that the author wrote confidently in April 2025 that DeepSeek had not distilled. Now, in light of the reasoning-trace-extraction research since, the author says it looks more possible than previously credited that DeepSeek did distill some o1 traces to help train R1, while stressing that this would not diminish R1's own innovation.
Key facts
- Interconnects published a curated reading list on open-weight AI models, last updated 13 September 2026, organized into three sections: Foundation, US-China Competition, and Technical Details.
- It puts the current open-closed AI performance gap at roughly four to six months and says the leading open models have all come from Chinese labs since about 2024.
- It cites a September 2026 Anthropic report, 'Detecting and countering misuse of AI,' describing at-scale use of Anthropic's products by banned parties that combined distillation via SFT data with undisclosed routing of Claude into their own services.
- It says lawmakers have probed DoorDash, Airbnb, Anysphere/Cursor and Apple over their use of Chinese AI models, while Perplexity adopted DeepSeek R1 and Thomson Reuters built on Qwen to move off Claude, a shift the author frames as cost-driven.
- Revisiting the 2025 debate over whether DeepSeek R1 was distilled from OpenAI's o1, the author now says it looks more possible than previously credited that DeepSeek did distill some o1 traces to help train R1, while stressing this does not diminish R1's own innovation.
Why it matters
This is not a single news event but a study guide, and it matters for what it stitches together in one place: business strategy, the safety debate and hard technical detail on a fast-moving question, explicitly framed as preparation for policy-facing writing. Its central technical claim, that the open-closed performance gap now sits at roughly four to six months and that Chinese labs have held the open-weight frontier since about 2024, is the backdrop against which the rest of the debate over whether to build on open models, whether to restrict them, and how much distillation actually explains China's position is playing out.
Who it affects
Enterprise teams deciding whether to build agentic workflows on open or closed models sit at the center of the list's own argument. So do policymakers weighing restrictions on open weights or on the use of Chinese models, and the four companies it says lawmakers have already probed over that exact question: DoorDash, Airbnb, Anysphere/Cursor and Apple. The labs whose disputes the list surveys, including OpenAI, Anthropic and DeepSeek, are also directly implicated, particularly in the distillation section's account of the R1/o1 dispute; unlike that dispute, its Anthropic misuse findings name no specific labs.
How to use it
The list is free to read on Interconnects, split into the three named sections so a reader can go straight to strategy, geopolitics or the technical distillation debate rather than reading start to finish. The author frames it as a standing resource rather than a one-off explainer and says it will keep being updated as readers suggest additions in the comments, so it works better as a bookmarked reference than a single sitting's read.
How solid is it
The list makes no new empirical claim of its own; its credibility rests on what it points to, including named analyst work from SemiAnalysis, Epoch AI and Artificial Analysis, an independent researcher's own gap analysis, a cited paper on reasoning-trace extraction, and mainstream reporting from Bloomberg, Reuters, CNBC, Forbes, Business Insider and Semafor for the regulatory and adoption claims. A large share of the entries, though, are the author's own earlier pieces, credited throughout to 'Nathan Lambert / Interconnects,' which makes the framing of the distillation debate closer to one recurring voice's argument than a survey of competing ones. That voice also revises itself within this same post: having written confidently in April 2025 that DeepSeek R1 was not distilled from OpenAI's o1, it now says newer research makes it look more possible than previously credited that DeepSeek did distill some o1 traces to help train R1, a useful sign that even someone this embedded in the debate is still updating.
Risks and caveats
It gives no methodology or benchmark behind the four-to-six-month gap figure, and it does not say which Chinese labs Anthropic's report identifies as having used the reasoning-trace-extraction technique, or what came of the lawmakers' probes into the four named companies. Since the author explicitly invites more suggestions and says the list will be updated over time, it should be read as a moving snapshot rather than a settled reference.
“Get it out fast. We open source it within a few hours.”
— Z.ai's product lead, quoted in 'The Z.ai Playbook,' ChinaTalk (Nov. 21, 2025)