Your Daily Briefs

DeepMind's Gemini-2.0 World Model — new paper on autonomous drone navigation using learned world models. Key insight: 40% fewer training samples vs prior art.

Anthropic's Constitutional AI v2 — updated framework with better harmlessness transfers across RLHF iterations.

HN Top: Elon Musk vs OpenAI — court dismissal signals stronger Big Tech AI enforcement. 829 points.

Google JAX + Pathways: new 100B world model trained on 32 TPU pods. 3x throughput vs prior architecture.

MIRIX's new interpretability paper: superposition hypothesis confirmed in transformer MLP layers.

OpenAI GPT-5 leaked benchmarks: 94.7% on MMLU, new reasoning benchmark exceeds human expert average.