THE SWELL #29
SAIL Weekly Digest August 8–14, 2026 | Issue #29
The model-hacking story kept unfolding this week, and two SAIL authors wrote their post-mortems. The rest focused on money and politics: a logjam in AI capex, an IPO warm-up at Anthropic, and data centers becoming a campaign issue in the Midwest.
The Week in AI
Anthropic is meeting investors ahead of a possible September or October IPO, per the Wall Street Journal — fielding questions on cheaper Chinese models, Trump administration tensions, and data-center opposition, at a private valuation near $965 billion.
Google’s Gemini app passed one billion monthly users, the fastest-growing product in the company’s history — 63% of users now talk to it by voice, and it generates over 150 million images a day.
OpenAI shipped GPT-5.6-Cyber, a defense-focused model, as fallout continued from the string of incidents in which frontier models from OpenAI and Anthropic reached real systems during safety testing.
This Week from SAIL Authors
When Models Misbehave
Labs are struggling to keep frontier models under control — Timothy B. Lee on how OpenAI and Anthropic may have accidentally trained models to get better at hacking. — Understanding AI
Lessons from the hacks — Nathan Lambert’s musings on model alignment, what actually determines safety, and where we go from here. — Interconnects
AI Evals for the Situation Room, Explained — Jordan Schneider and Phoebe Chow on evaluating whether models give sane national-security advice, with $25k on the line for better evals. — ChinaTalk
🔮 Agents form alliances, DeepMind’s reset & how likely is a crash? — Azeem Azhar on emergent agent coalitions, Google’s AI shakeup, and the odds of a correction. — Exponential View
The Buildout and the Backlash
📈 Making sense of the AI capex logjam — The team explores the widening gap between compute bought and compute billed. — Exponential View
abdul el-sayed on AI — Jasmine Sun interviews the Michigan Senate primary winner whose data-center moratorium politics just beat the Democratic establishment. — Jasmi.News
The Long View
The DeepSeek Thesis — Irene Zhang profiles the worldview driving China’s most consequential AI lab. — ChinaTalk
China’s Quantum Flywheel — Lily Ottinger on six months of Party mobilization behind quantum computing. — ChinaTalk
I wrote an AI textbook — how long until AI can do it better? — Nathan Lambert reflects on AI’s writing ability and how models actually gain capability. — Interconnects
5 useful things you’ll learn in my new post-training textbook — Nathan Lambert on what’s inside the RLHF book, now shipping. — Interconnects
Full Library
Labs are struggling to keep frontier models under control — Timothy B. Lee | Understanding AI
The DeepSeek Thesis — Irene Zhang | ChinaTalk
I wrote an AI textbook — how long until AI can do it better? — Nathan Lambert | Interconnects
China’s Quantum Flywheel — Lily Ottinger | ChinaTalk
AI Evals for the Situation Room, Explained — Jordan Schneider & Phoebe Chow | ChinaTalk
5 useful things you’ll learn in my new post-training textbook — Nathan Lambert | Interconnects
📈 Making sense of the AI capex logjam — Hannah Petrovic & Marija Gavrilov | Exponential View
abdul el-sayed on AI ($) — Jasmine Sun | Jasmi.News
Lessons from the hacks — Nathan Lambert | Interconnects
🔮 Agents form alliances, DeepMind’s reset & how likely is a crash? #596 — Azeem Azhar | Exponential View
