Bliss Reading Group - June 8
Reading group on video generation, world models, and physical generalization.
The BLISS Reading Group will take place on June 8, 2026 at the Merantix AI Campus in Berlin. This session explores whether modern video generation models actually learn underlying physical laws or mainly produce convincing predictions within familiar data distributions.
The discussion centers on the paper “How Far is Video Generation from World Model: A Physical Law Perspective” by Kang et al. (2024). Inspired by models such as Sora, which generate videos that appear to follow physics, the paper tests whether video generation models can truly generalize. Its findings suggest that while models perform well in-distribution and show some progress on combinatorial tasks, naive scaling alone is not enough to discover physical laws.
The session is designed for researchers, AI practitioners, students, and anyone interested in video generation, world models, and the limits of current AI architectures. The discussion will focus on what is needed for models to extrapolate beyond training data rather than simply interpolate.
- Date: June 8, 2026
- Time: 6:45–8:00 PM
- Location: Merantix AI Campus, Max-Urich-Straße 3, 13355 Berlin
- Organizer: BLISS Berlin and Merantix AI Campus
- Admission: Approval required
- Audience: AI researchers, ML practitioners, students, world model enthusiasts
- Focus: Video generation, world models, physical laws, generalization