Curated internet, served daily
Models, agents, and what actually runs
Models, agents, and the tools built on top of them — what runs, on whose hardware, at what cost. I care about the gap between a benchmark and a Tuesday, so most of what lands here is something I ran rather than something I read about.
Two old bug classes — a heap overflow and an SSO misconfiguration — carried a paid security team into OpenAI's internal repositories in under 72 hours, and the identity mistake is the part that decided how far the breach could travel.
PrismML's ternary 27B multimodal model lands at 5.9GB with 98.2% of full-precision benchmark performance, and the retention number is the part worth checking.
Tencent's BrowserSkill hands any shell-capable agent the browser you're already signed into, and picked up 1,350 stars in a single day.
TypeSafe AI's Jev trades text generation for calibrated, typed decisions, claiming two orders of magnitude more speed and a model that can't hallucinate.
A pure-C engine with zero dependencies promises frontier MoE models on hardware you already own, and picked up 2,233 stars in a single day.
VoiceStudio claims a fully local voice stack with cloning, dubbing and transcription in 646 languages — and a star count on the same page that contradicts itself.