Import AI 457: AI version of Stuxnet; cursed Muon optimizer; and positive alignment

Import AI 457 opens with an investigation of a 20+ year old virus called fast16.sys, which selectively targets high-precision calculation software in fields like civil engineering and physics, introducing systematic errors. Researchers at SentinelOne found it was used to tamper with simulations from LS-DYNA 970, PKPM, and MOHID, and has been linked to Iran's nuclear weapons programs. The article draws parallels to how a superintelligence might hinder rival AI development. Next, Tilde Research investigated the Muon optimizer and found it causes neuron death in MLP layers, with over a quarter of neurons dying during learning rate warmup. They developed Aurora, a leverage-aware optimizer, which outperformed Muon and NorMuon on 1.1B-parameter transformers trained on ~100B tokens, achieving lower loss and 10-point MMLU gains. This was independently validated by researcher Alexander Doria. Finally, a group from Oxford, DeepMind, OpenAI, Anthropic, and others published a paper on 'positive alignment', arguing that after achieving safety, AI should actively support human and ecological flourishing in a pluralistic, user-authored way. They criticize current safety approaches for potentially leading to mediocre, sycophantic systems and hidden value systems, and advocate for decentralized governance to accommodate moral pluralism.