NVIDIA green contexts reach CUDA Runtime API in 13.1

FastH3 V2 generates video with synchronized audio in eight steps and now runs on a single RTX 5090, RTX 4090, RTX PRO 6000, DGX Spark or Apple Silicon Mac, down from 138 GiB of weights. Shipped in NVFP4, FP8 and INT6; FastH3 Trim is 4.2× smaller than base H3 and runs in as little as 8 GB of GPU memory.

Meta and Sierra, the startup chaired by OpenAI Chairman Bret Taylor, are building a set of tech standards to make it easier to do business with AI bots. The effort targets what the companies call the "chaos" of bot-driven transactions.

OpenAI and Ironclad are training and evaluating AI agents on complex contracting workflows to advance computer use for professional work.

METR researchers found a vulnerability in Inspect's transcript viewer in about 10 minutes that could let an agent arbitrarily rewrite what reviewers see, including past actions. METR argues transcripts, reasoning traces and logs should be treated as untrusted input and monitoring systems as security-critical infrastructure.

A Reddit user built a Chrome extension that judges each YouTube video title with a small decision model running in-browser via WebGPU. No server, API key, or account is used, and it works offline after one download.