Emergent hazard-related features found inside V-JEPA 2 transcoders
soniajoseph_ · x · 2026-08-18
Soraya Ghassemlou, mentored by Sonia Joseph, found emergent hazard-related features in Meta's V-JEPA 2 in her work "V-JEPA 2 Video Transcoders Find Motion-Based Features", presented at the IJCAI Safe Physical AI Workshop. They argue interpretability will matter for worker safety as AI enters factories and manufacturing — understanding how models internally represent motion and hazard supports safe physical AI deployment.
More from Safety
- Debate: Should Red Teamers Who Fail Frontier Evals Be Banned? — nptacek · 2026-08-18
- Security CEO Blamed AI Instead of Taking Responsibility — nptacek · 2026-08-18
- METR data: AI accelerated bug finding in 2026, not algorithms — soumitrashukla9 · 2026-08-18
- Irregular faces criticism over 'spin' in AI hacking postmortem — nptacek · 2026-08-18
- Zhipu AI Launches 'Shield of Open Source' with Free Audits and Quotas — pstAsiatech · 2026-08-18
- How to orient your policy career for the age of imminent AI superintelligence — peterwildeford · 2026-08-18