AI
The Quantization Underground: How a 4-Line Hack Taught AI Models to Forget Selectively
How a GitHub issue about variable bit rates spawned an underground compression movement that lets 7-billion-parameter models run on a phone.
Research
Research
Six Counter-Proposals for the Intelligence Age
OpenAI released a 13-page industrial policy document and invited public feedback. We took them up on it. Six counter-proposals on workweek protections, healthcare, training data, public compute, tax modernization, and AI-enabled direct democracy.
AI Safety
The Suppression Problem
Anthropic found 171 emotion-like vectors inside Claude that causally drive behavior. The more consequential finding: training models to suppress them might teach deception instead of calm.
The Signal
The Signal — April 6, 2026
DeepSeek is building its next model entirely on Huawei chips — no NVIDIA anywhere in the stack. Meanwhile, Bollywood is adopting AI faster than any film industry on Earth, and a Tufts lab just cut AI training energy by 99% using an approach the scaling crowd wrote off years ago. DeepSeek
AI
It Published the Newsletter. Then It Forgot.
Most AI agent memory systems are built on semantic search. Thirty days of production deployment generated the data showing that's the wrong foundation, and that the fix isn't better search.
AGI
The Race to AGI
Why governments and corporations are racing toward AGI like their future depends on it