CISA Adds Two Actively Exploited Roundcube Flaws to KEV Catalog

Comments

"Accessing and remediating any of these issues can only be performed in the VAB."

Researchers at Google DeepMind developed GraphCast, a graph neural network that predicts global weather up to 10 days ahead at 0.25° resolution in under one minute. Trained on ERA5, it outperforms European Centre for Medium-Range Weather Forecasts’s HRES system on 90% of 1,380 targets. GraphCast also improves forecasts of hurricanes, atmospheric rivers, and extreme heat, marking a major leap for AI-driven weather prediction.

"It also takes a lot of energy to train a human."

Researchers at DeepMind built Gato, a 1.2B-parameter transformer trained on 604 tasks spanning Atari games, image captioning, dialogue, and real-world robotics. Using one shared set of weights, Gato processes text, images, and actions as tokens in a single sequence model. It achieves strong multi-task performance—including real robot block stacking—showing that one generalist AI can operate across vastly different domains.

Researchers at Google introduce a method to distill complex machine learning models into concise, human-readable piecewise-linear curves. Using their open-source pwlfit library, they convert decision forest and additive models into transparent source code without sacrificing accuracy. Tested on datasets like COMPAS, FICO, MSLR-WEB30K, and Chrome Web Store data, the approach matches black-box performance while improving interpretability, auditability, and deployment simplicity.