Gemini Omni 1.2 Flash Brings 15-Second Video

Gemini Omni 1.2 Flash Adds 15-Second Video Generation

Google made Gemini Omni 1.2 Flash generally available, adding native 15-second video generation with stronger character and brand consistency plus multi-turn conversational editing. The price also dropped, making video generation cheaper at scale. — blog.google

CEIA's Lumi-4B Puts Brazilian Portuguese First

Brazil's CEIA/UFG research team released Lumi-4B, a small open language model tuned for Brazilian Portuguese. Continual pre-training and distillation from a larger teacher model give it Portuguese fluency while keeping its math and coding strengths. — huggingface.co

New Scale Grades AI Labs on Safety Culture

A new framework grades AI developers on a safety-culture scale from -2 to +7, measuring how fast they detect and fix dangers rather than static odds of doom. Leading labs currently rate near the bottom, far below the level needed for superintelligence to go well. — lesswrong.com

AI Bots Swamp Academic Library Websites

AI data-scraping bots are flooding academic library websites, slowing or crashing collections that students and researchers rely on; one university saw 11,000 bot searches in a single day. Major AI firms now support a standard opt-out signal that lets sites block their crawlers, although compliance remains voluntary. — cbc.ca

Ripple-65B Preview Now Runs Locally on llama.cpp

The first local-ready versions of Ricursive's Ripple-65B Preview are out. The open-weights coding model now runs in llama.cpp, the standard open-source runtime for running models on your own hardware, with support merged upstream. — huggingface.co