SentinelOne

Import AI 457: Stuxnet-like AI, cursed Muon optimizer, and positive alignment

SentinelOne researchers discovered a 20-year-old computer virus called fast16 that selectively distorts results of high-precision calculations, similar to futuristic weapons against scientific progress. Tilde Research identified a serious flaw in the Muon optimizer—it kills neurons in MLP layers—and proposed an alternative, Aurora. A group of scientists introduced the concept of 'positive alignment' for AI, aiming not only for safety but also for actively promoting human flourishing.

SentinelOne Prime Intellect OpenAI Anthropic Google DeepMind