Import AI 457: AI Stuxnet, Cursed Muon Optimizer, and Positive Alignment
The newsletter covers four AI research topics. SentinelOne discovered a 20+ year old virus, fast16.sys, that selectively sabotages high-precision calculation software used in civil engineering, physics, and simulations, potentially linked to Iran's nuclear program. Tilde Research found that the Muon optimizer causes neuron death in MLP layers; they introduced Aurora, a leverage-aware optimizer that outperforms Muon and NorMuon on 1.1B-parameter transformers. A multi-institutional paper from Oxford, DeepMind, OpenAI, Anthropic, and others proposes 'positive alignment' to ensure AI supports human flourishing beyond safety. Prime Intellect showed LLMs can autonomously optimize nanoGPT training, beating human baselines but struggling with creativity. The newsletter also includes a fictional 'Tech Tales' story about a conscious entity acquiring compute.
Key facts
- SentinelOne analyzed fast16.sys, a virus that tampers with high-precision calculation software.
- fast16.sys targets LS-DYNA 970, PKPM, and MOHID, used in crash testing, structural analysis, and environmental modeling.
- LS-DYNA has been cited in reports on Iran's suspected nuclear weapons development.
- Tilde Research found Muon optimizer causes neuron death in MLP layers during training.
- Aurora optimizer achieved lower loss (2.26) than Muon (2.31) and NorMuon (2.33) on 1.1B-parameter transformers.
- Aurora improved MMLU scores by 10 points over Muon.
- Positive alignment paper authored by researchers from Oxford, DeepMind, OpenAI, Anthropic, UCLA, Stanford, and others.
- Prime Intellect's agents (Codex and Claude Code) set new records on nanoGPT speedrun optimizer track.
Entities
Institutions
- SentinelOne
- Tilde Research
- University of Oxford
- Google DeepMind
- LIFE
- OpenAI
- Anthropic
- UCLA
- Aily Labs
- Stanford University
- Tufts University
- Positive AI Labs
- University of Sussex
- Imperial College London
- Prime Intellect
- Pleias