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

Analysts at SentinelOne analyzed the 20-year-old computer virus fast16.sys and found that it selectively attacked high-precision computational software, modifying code in memory to distort computation results. The virus targeted engineering and simulation packages such as LS-DYNA 970, PKPM, and MOHID, which are also used in nuclear weapons modeling. Researchers suggest this could have been part of a program to slow scientific progress in certain countries. In another study, Tilde Research discovered that the Muon optimizer causes neuron death in MLP layers due to anisotropy of row norms, degrading model quality. In response, they developed the Aurora optimizer, which accounts for matrix "leverages." On models with 1.1 billion parameters trained on ~100 billion tokens, Aurora achieved lower loss and improved benchmark results, including a 10-point gain on MMLU compared to Muon. A group of researchers from Oxford, DeepMind, OpenAI, Anthropic, and other organizations published a position paper introducing the concept of "positive alignment." Unlike the traditional approach focused on preventing failures, positive alignment involves creating AI systems that actively support human and ecological flourishing, accounting for value pluralism and decentralized governance. Additionally, Prime Intellect conducted an experiment where LLM agents (Codex and Claude Code) autonomously optimized the training of another model, nanoGPT, performing about 10,000 runs and consuming ~14,000 H200 hours. The agents surpassed the human baseline but failed to generate novel ideas and tended to endlessly accumulate components without refactoring them.