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

Researchers at SentinelOne discovered a computer virus over 20 years old named fast16.sys that 'selectively targets high-precision calculation software, patching code in memory to tamper with results.' The virus uses a self-propagation mechanism designed to 'produce equivalent inaccurate calculations throughout a facility.'

The injected code block contains 'a larger and complex sequence of floating-point unit instructions specifically for precision arithmetic and scaling values in internal arrays,' which is a 'standalone mathematical calculation function unrelated to code flow hijacking or any other typical malicious code injection.'

After converting the patch rules into YARA signatures, researchers matched 'fewer than ten files' in a large corpus, which 'share a clear theme: they are precision calculation tools in specialized fields such as civil engineering, physics, and physical process simulation.' The strongest overlaps point to three high-precision engineering and simulation suites: LS-DYNA 970, PKPM, and the MOHID hydrodynamic modeling platform. LS-DYNA has been publicly reported in connection with Iran's alleged violation of JCPOA Section T for computer modeling research related to nuclear weapons development.

Researchers at Tilde Research discovered a serious flaw in the Muon optimizer: 'Muon's updates inherit row-norm anisotropy on tall matrices, which can cause a significant portion of neurons in MLP layers to permanently die.' Specifically, 'under Muon, neurons are initially active with uniformly high leverage, but during learning rate warm-up, a large number of neurons die and cannot recover. By step 500, more than a quarter of neurons are effectively dead.'

In response, researchers proposed Aurora, a 'lever-aware optimizer for rectangular matrices.' After training a 1.1B parameter transformer model on approximately 100B tokens, 'Aurora achieved the lowest final loss among all methods, with a smooth loss of 2.26 at step 24k, significantly outperforming Muon (2.31) and NorMuon (2.33).' The loss improvement translated into consistent gains on standard benchmarks: 'Aurora improved MMLU scores by 10 points.' Pleias researcher Alexander Doria has independently validated Aurora's superiority over Muon and AdamW on a 600M parameter model.

Researchers from Oxford University, Google DeepMind, OpenAI, Anthropic, and other institutions published a position paper introducing the concept of 'Positive Alignment.' They define it as 'developing AI systems that (i) remain safe and cooperative, and (ii) actively support human and ecological flourishing in a pluralistic, polycentric, context-sensitive, and user-empowering manner.'

The paper states: 'Over the past decade, negative alignment has rightly prioritized failure mode reduction. However, if we want AI systems to improve human outcomes in the contexts where they are actually used, we may need an additional research program that treats alignment as constructive support for human goals.' They warn that 'adopting only a negative posture may optimize our information ecology for risk aversion rather than human development, potentially reducing catastrophic errors while trapping society in a local optimum of superficial and 'soulless' assistance.'