Import AI 457: AI version of Stuxnet; cursed Muon optimizer; and positive alignment
This issue of Import AI covers three stories: an ancient computer virus 'fast16' that sabotages high-precision calculations, reminiscent of a superintelligence's non-proliferation tactics; the discovery that the Muon optimizer can kill neurons in MLP layers, leading to the development of a replacement called Aurora; and a position paper advocating for 'positive alignment' to ensure AI supports human flourishing beyond mere safety.
SentinelOne
Prime Intellect
OpenAI
Anthropic
Google DeepMind
Import AI27.07 · 12:06
