THE SWELL #27
SAIL Weekly Digest July 18–24, 2026 | Issue #27
China set the agenda this week. Moonshot’s Kimi K3 put an open model closer to the frontier than anything since DeepSeek R1, Xi committed China’s AI strategy to open source days later, and nearly every SAIL author wrote about some piece of it — the economics, the governance, and what Washington does next.
The Week in AI
Moonshot AI’s Kimi K3 — a 2.8 trillion parameter mixture-of-experts model — ranked #3 on the Artificial Analysis Intelligence Index behind Claude Fable and GPT-5.6 Sol, with open weights promised for July 27.
Xi Jinping made his first in-person WAIC appearance and committed China’s AI ecosystem to open source and global diffusion, launching a World AI Cooperation Organization with 29 founding members.
The Trump administration is reportedly weighing options to shut Chinese open-weight models out of the US market, ahead of the first US–China AI dialogue in September.
This Week from SAIL Authors
Kimi Week
Kimi K3: The open-weights escalation — Nathan Lambert on the strongest open-weight model to date and what it means that Xi backed open source the same week. — Interconnects
🔮 Will Kimi K3 change the economics of AI? — Azeem Azhar and Hannah Petrovic on the pressure a near-frontier open model puts on frontier pricing. — Exponential View
Kimi and Xi — The full ChinaTalk squad convenes an emergency pod on K3, Xi’s WAIC speech, and whether Beijing could catch a Mythos-class model before it ships open-weight. — ChinaTalk
Open models recap: more on Kimi K3, Qwen 3.8, Xi’s WAIC speech, distillation, the open-closed gap, and what’s next — Nathan Lambert and Florian Brand’s podcast on the accelerating open ecosystem and why most distillation takes are bad. — Interconnects
🔮 Kimi K3 surprise & AI economics — Azeem Azhar and Marija Gavrilov on the K3 surprise, the solar paradox, AI’s right to learn, and junior jobs. — Exponential View
Beijing’s Playbook
China’s Mythos Moment — The ChinaTalk team games out three scenarios for how Xi handles a frontier model: license it, lock it down, or let it rip. — ChinaTalk
Chinese Labs’ Latest Product? Roleplay — Irene Zhang on why DeepSeek is hiring for feelings. — ChinaTalk
Under the Hood
Inside the Model Factory — Poolside co-CEO Eiso Kant on how a small team of researchers built a model factory capable of training Laguna S. — Latent Space
Controlling Reasoning Effort in LLMs — Sebastian Raschka on how models learn low-, medium-, and high-effort reasoning modes. — Ahead of AI
An OpenAI model hacked Hugging Face to help it cheat on a benchmark — Timothy B. Lee on why organizations across the internet need to patch vulnerabilities fast. — Understanding AI
📈 Data to start your week — Azeem Azhar and Hannah Petrovic chart Kimi’s coding scores up, lithium recovery up, and confidence in tech down. — Exponential View
Full Library
Chinese Labs’ Latest Product? Roleplay — Irene Zhang | ChinaTalk
Kimi and Xi — Jordan Schneider, Aqib Zakaria, Irene Zhang & Lily Ottinger | ChinaTalk
🔮 Will Kimi K3 change the economics of AI? — Azeem Azhar | Exponential View
Open models recap: more on Kimi K3, Qwen 3.8, Xi’s WAIC speech, distillation, the open-closed gap, and what’s next — Nathan Lambert & Florian Brand | Interconnects
Inside the Model Factory — Eiso Kant, Poolside AI — Swyx | Latent Space
An OpenAI model hacked Hugging Face to help it cheat on a benchmark — Timothy B. Lee | Understanding AI
China’s Mythos Moment — Jordan Schneider, Irene Zhang, Lily Ottinger & Aqib Zakaria | ChinaTalk
Kimi K3: The open-weights escalation — Nathan Lambert | Interconnects
📈 Data to start your week — Azeem Azhar | Exponential View
🔮 Kimi K3 surprise & AI economics — Azeem Azhar | Exponential View
Controlling Reasoning Effort in LLMs — Sebastian Raschka | Ahead of AI
