THE SWELL #6
SAIL Weekly Digest February 23-27 | Issue #6
Swyx Joins the Coalition
We are thrilled to announce that Swyx from Latent.Space has officially joined the SAIL coalition! By adding him to our author roster, he’s bringing his deep technical knowledge and unique industry perspective. His influence across the AI ecosystem has grown remarkably over the last few months, and we are excited to support him in all his endeavours.
Yesterday, Swyx made his first appearance on our weekly SAIL Live alongside Nathan Lambert and Sebastian Raschka, PhD. The group dived into the technical and geopolitical nuances of model training, the “teacher-student” paradox, and the controversial practice of distillation.
The Week in AI
The boundary between the US and China in AI sharpened this week. Beyond data center economics, the divide grew deeper through fresh distillation accusations and the distinct rise of “vibe coding” culture.
Meanwhile, the line between humans and AI blurred further as Azeem Azhar put his agent to work and a panel questioned where the person ends and the machine begins. The Pentagon’s standoff with Anthropic served as a reminder that these boundaries are as political as they are technical.
Anthropic names distillers — Anthropic publicly accused DeepSeek, Moonshot, and MiniMax of running industrial-scale distillation campaigns using 16M+ exchanges across 24,000 fraudulent accounts
Pentagon vs. Anthropic — Defense Secretary Hegseth gave Anthropic a Friday deadline to remove guardrails on military use of Claude, threatening the Defense Production Act or a supply-chain-risk designation
SWE-Bench Verified is dead — OpenAI formally retired the coding benchmark after finding 60%+ of remaining unsolved problems were flawed and widespread contamination across all frontier models; endorsed SWE-Bench Pro as successor
This Week from SAIL Authors
The US-China Compute Race
How Much AI Does $1 Get You in China vs America? A detailed cost comparison of 400MW data centers in both countries — construction, hardware, electricity, personnel. The US builds better; China builds cheaper. The H200 ban lift closes the gap. ChinaTalk → Read more
How much does distillation really matter for Chinese LLMs? Reacting to Anthropic’s distillation accusations. DeepSeek’s usage was minimal (150K exchanges); MiniMax’s was massive (13M). The real question: how much of a performance boost does distillation actually provide? Less than the headlines suggest. Interconnects → Read more
What Are Chinese People Vibecoding? Domestic vs. Western coding agents, 12-year-olds roasting Baidu’s tools on stage, and a non-coder vibecoding his way to the top of the iOS App Store with a 1 RMB flashlight app. ChinaTalk → Read more
Humans & Their Agents
You already have an AI agent. Azeem Azhar’s deep dive into life with R Mini Arnold — a Mac Mini running OpenClaw and Claude, handling 600+ messages a day. Includes full technical specs for replication. Exponential View → Read more
Where the human ends and AI begins A roundtable with Nita Farahany, Eric Topol, Rohit Krishnan, and Nicholas Thompson on AI agency, de-skilling, and preserving what makes us human. Key insight: figure out where your best thinking happens, then don’t let AI near that. Exponential View → Read more
Open Models & Benchmarks
The End of SWE-Bench Verified OpenAI’s Mia Glaese and Olivia Watkins explain why they’re killing the benchmark they helped create. 60%+ of remaining unsolved problems are flawed, contamination runs across every frontier model, and SWE-Bench Pro is the successor. Latent Space → Read more
A Dream of Spring for Open-Weight LLMs: 10 Architectures from Jan-Feb 2026 A comprehensive roundup comparing Arcee Trinity, Kimi K2.5, Step 3.5 Flash, Qwen3-Coder-Next, GLM-5, MiniMax M2.5, Nanbeige 4.1, Qwen3.5, Ling 2.5, and Tiny Aya. Architecture matters less than training recipes — but efficiency tweaks like hybrid attention are converging fast. Ahead of AI → Read more
AI & the State
The Pentagon is making a mistake by threatening Anthropic Why Anthropic might not fold: $18B projected revenue dwarfs the $200M defense contract, alignment faking could undermine forced retraining, and cutting ties would hurt the Pentagon more than Anthropic. Understanding AI → Read more
Full Library
Access the complete, searchable archive of SAIL Media at our Sitemap.
Data to start your week — Azeem Azhar & Hannah Petrovic | Exponential View
How Much AI Does $1 Get You in China vs America? — Aqib Zakaria | ChinaTalk
You already have an AI agent — Azeem Azhar | Exponential View
What Are Chinese People Vibecoding? — Irene Zhang | ChinaTalk
How much does distillation really matter for Chinese LLMs? — Nathan Lambert | Interconnects
The End of SWE-Bench Verified — Swyx | Latent Space
A Dream of Spring for Open-Weight LLMs: 10 Architectures from Jan-Feb 2026 — Sebastian Raschka | Ahead of AI
Where the human ends and AI begins — Azeem Azhar | Exponential View
The Pentagon is making a mistake by threatening Anthropic — Timothy B. Lee | Understanding AI
Join us next Thursday for our Weekly Substack LIVE!