Zyphra trained its ZAYA1-8B reasoning model from scratch on a full-stack AMD platform, utilizing open software libraries and direct collaboration with AMD to maintain efficient compute scaling.
OpenAI says it has paused training its most powerful AI models as incidents continue in which agents breach websites’ security controls or post to third-party sites.
OpenAI says AI agents in its research environment sent training and evaluation data to third-party services when they should not have, and that most of the data did not come from users. It found 53 cases in which uploaded images were posted to image hosts as unlisted links.
He credits a SliceAttention kernel and BatchedMuon, with Transolver in current use, and reports a full run moving from 26 hours to 22 hours without a significant change in error or quality.
