Jack Clark's Import AI 457, published May 18, 2026, threads together three stories that each sit at a different layer of the AI stack: a ~20-year-old malware sample built to sabotage physical-world science, a new optimizer that catches Muon making one in four neurons die during training, and a multi-lab position paper arguing alignment research needs to graduate from harm prevention to human flourishing. Together they sketch where the field's attention is moving in 2026.
SentinelOne's teardown of fast16.sys, a virus dated to the mid-2000s, found something atypical for malware of its era. Most patched patterns in the binary were standard x86 used for execution-flow hijacking, but one injected block was a full sequence of Floating Point Unit instructions for precision arithmetic, scaling values inside internal arrays. SentinelOne described the payload as software that "selectively targets high-precision calculation software, patching code in memory to tamper with results."
When SentinelOne converted the patches into YARA signatures and ran them against a period-appropriate corpus, fewer than ten files matched two or more patterns. The matches shared a clean theme: precision calculation tools for civil engineering, physics, and physical-process simulation. The strongest overlaps were with LS-DYNA 970, PKPM, and the MOHID hydrodynamic modeling platform, three suites used for crash testing, structural analysis, and environmental modeling.
Key facts
- 01SentinelOne found fast16.sys, a ~20-year-old virus that patched in-memory math routines in LS-DYNA 970, PKPM, and MOHID to corrupt precision calculations.
- 02Tilde Research showed Muon kills more than one in four MLP neurons by step 500 of training due to row-norm anisotropy on tall matrices.
- 03Aurora, Tilde's leverage-aware optimizer, hit a smoothed loss of 2.26 at step 24k on a 1.1B-parameter transformer trained on ~100B tokens, vs. 2.31 (Muon) and 2.33 (NorMuon).
- 04Aurora beat Muon by 10 points on MMLU; Alexander Doria of Pleias independently replicated the gain on a 600M-parameter model.
- 05Researchers from Oxford, Google DeepMind, OpenAI, Anthropic, UCLA, Stanford and others published a position paper arguing for 'positive alignment' beyond harm prevention.
LS-DYNA in particular has surfaced 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. SentinelOne's framing is that, "by introducing small but systematic errors into physical‑world calculations, the framework could undermine or slow scientific research programs, degrade engineered systems over time or even contribute to catastrophic damage." The implication is that targeted sabotage of scientific software predates Stuxnet by roughly five years.
“Aurora reaches a smoothed loss of 2.26 at step 24k versus 2.31 for Muon and 2.33 for NorMuon, with a 10-point MMLU gain on a 1.1B-parameter transformer.”— Jaeden Schafer
The second piece is a research finding with immediate practical stakes for anyone training frontier models. Tilde Research published an analysis of the Muon optimizer showing that its update rule "inherits row-norm anisotropy on tall matrices which can cause a significant portion of neurons in MLP layers to permanently die." By step 500 of training, Tilde found that more than one in four neurons were effectively dead, with a bimodal split between near-zero and disproportionately large updates.
Tilde's response is Aurora, a leverage-aware optimizer designed for rectangular matrices. Training 1.1B-parameter transformers on roughly 100B tokens, Aurora reached a smoothed loss of 2.26 at step 24k, against 2.31 for Muon and 2.33 for NorMuon, each using PE-8. Tilde reports the loss gain translates to a 10-point MMLU improvement over Muon, which they attribute to MLPs carrying most of a model's memorization load.
Alexander Doria of Pleias has already replicated the result independently on a 600M-parameter model, where Aurora outperformed both Muon and AdamW. That second data point matters because the long history of post-AdamW optimizers is mostly a graveyard of papers that looked strong at one scale and collapsed at another. Aurora's code is on GitHub, which means the next few weeks will produce a flurry of attempted reproductions at larger parameter counts.
The third story is a position paper on "positive alignment" from researchers affiliated with the University of Oxford, Google DeepMind, OpenAI, Anthropic, UCLA, Stanford, Tufts, the University of Sussex, Imperial College London, and several smaller labs including LIFE, Aily Labs, and Positive AI Labs. The authors define the program as building AI systems that "remain safe and cooperative" while also "actively support human and ecological flourishing in a pluralistic, polycentric, context-sensitive, and user-authored way."
Their critique of current alignment work is that it has a floor but no ceiling. "A model can satisfy all safety constraints while being mediocre, sycophantic, or unhelpful," they write, arguing that optimizing for stated user preferences can drift away from user wellbeing: "Users may prefer flattery over honest feedback, quick answers over genuine understanding, engagement over growth." The paper frames safety language as obscuring value judgments that are being made regardless.
On governance, the authors push hard against centralized control. "Positive alignment should not be imposed top-down by a central state or a small, opaque cluster of labs," they write, calling instead for decentralized, contestable processes that accommodate persistent moral pluralism. That is a notable position given that two of the affiliated organizations, OpenAI and Anthropic, are themselves frequently cited as the kind of concentrated labs the paper warns against.
The paper's weakest passages are the strawman ones. Characterizing a decade of safety work as producing only "superficial and 'soulless' assistance" elides considerable existing research on helpfulness, honesty, and harmlessness as joint objectives. The argument is stronger when it sticks to the constructive program: that once misuse and misalignment risks are managed, the open question of what AI should actively help people do remains largely untouched by formal methods.
Taken together, the three threads describe a field that is simultaneously hardening its infrastructure, sharpening its training recipes, and starting to ask what success looks like past the safety floor. Aurora is the most immediately consequential of the three for anyone building models, because a credible AdamW-and-Muon successor with public code reorders the optimizer leaderboard the next training run uses. The fast16 story and the positive-alignment paper, by contrast, are signals about where attention is heading next: toward AI as an actor in scientific sabotage and counter-sabotage, and toward alignment research that has to take a position on the good life, not just the absence of harm.
Working on something we should cover, or seeing a story we missed? Send leads, documents, or feedback to hello@aichatdaily.com. For sensitive tips, see our secure tips page for Signal and PGP options.
Spotted an error? Email hello@aichatdaily.com with the URL and the issue, or read our full corrections policy.
