Import AI 457: AI stuxnet; cursed Muon optimizer; and positive alignment

SentinelOne researchers published an investigation of fast16.sys, a roughly 20-year-old computer virus they say selectively targets high-precision calculation software, patching code in memory to tamper with results and combining that payload with self-propagation mechanisms. They report that most patched patterns correspond to standard x86 code used for hijacking execution flow, but one injected block is a larger, complex sequence of Floating Point Unit instructions dedicated to precision arithmetic and scaling values in internal arrays. Converting the patching rules into hexadecimal YARA signatures and running them against a period-appropriate corpus produced fewer than ten matching files, which the researchers say shared a theme of precision calculation tools in domains such as civil engineering, physics and physical process simulations. The strongest overlaps, they write, point to LS-DYNA 970, PKPM and the MOHID hydrodynamic modeling platform from the mid-2000s, used for crash testing, structural analysis and environmental modeling. LS-DYNA has been cited in public reporting on Iran's suspected violations of Section T of the JCPOA and in studies of computer modeling relevant to nuclear weapons development, they note.

Tilde Research examined the Muon optimizer and reported that its update inherits row-norm anisotropy on tall matrices, which they say can cause a significant portion of neurons in MLP layers to permanently die. By step 500, they write, more than one in four neurons are effectively dead, producing a sharply bimodal distribution of leverage scores. They built Aurora, a leverage-aware optimizer for rectangular matrices, and trained 1.1B-parameter transformers on about 100B tokens, reporting a smoothed loss of 2.26 at step 24k versus 2.31 for Muon and 2.33 for NorMuon, with MMLU scores 10 points above Muon. Researcher Alexander Doria independently validated Aurora against Muon and AdamW on a 600M-parameter model.

A position paper on "positive alignment" involves authors affiliated with the University of Oxford, Google DeepMind, LIFE, OpenAI, Anthropic, UCLA, Aily Labs, Stanford, Tufts, Positive AI Labs, the University of Sussex and Imperial College London. It defines the term as developing AI systems that remain safe and cooperative and actively support human and ecological flourishing in a pluralistic, polycentric, context-sensitive, user-authored way, and argues positive alignment should be expressed through decentralized, contestable processes rather than imposed top-down.

Prime Intellect tested Codex (GPT 5.5) and Claude Code (Opus 4.7) on the nanoGPT speedrun optimizer track, which requires training a 124M-parameter GPT-style model. The agents did about 10k runs using roughly 14k H200 hours and beat the human baseline in every session, but struggled to generate new ideas independently and rarely pruned components, the organization said.