Veo 3: Emergent Zero‑Shot Video Intelligence Toward Vision Foundation Models

Veo 3 demonstrates wide-ranging zero-shot performance across perception, physical modeling, visual manipulation, and reasoning tasks. These capabilities arise without task-specific training, echoing the emergence seen in LLMs trained on web-scale data. The authors conclude that video models are poised to become generalist vision foundation models.

Key Points

Generative video models trained at scale exhibit emergent zero-shot capabilities across a wide spectrum of visual tasks.
Veo 3 performs perception tasks (from edge detection to interpreting classic visual illusions) without task-specific training.
It demonstrates physical modeling and world understanding (material properties, dynamics, optics, color mixing, memory).
It can manipulate and create visuals (editing, composition, 3D reposing, novel views) and perform dexterous, affordance-aware actions.
Early forms of visual reasoning (graph/maze solving, sequences, analogies, rule extrapolation) emerge, suggesting a path to vision foundation models.

Sentiment

The community is predominantly positive, impressed by the breadth and systematicness of Veo 3's demonstrated zero-shot capabilities. Most commenters see these results as compelling evidence for the trajectory toward unified vision foundation models. The main dissenting voice argues from a philosophy-of-mind perspective that these benchmarks are fundamentally misguided, but this position is not widely supported and the associated thread was partially flagged by the community.

In Agreement

The systematic study of video models' emergent zero-shot capabilities is a valuable contribution, building on ideas previously circulating informally
Video-trained models build superior spatial understanding that transfers even to still image generation, outperforming dedicated image models for tasks like human anatomy
AI training can be reframed as teaching models to extrapolate future states from a starting condition plus intent, which naturally yields broad problem-solving ability
These findings suggest convergence toward a single unified model architecture capable of general-purpose vision tasks

Opposed

The benchmark categories used (segmentation, affordance recognition, etc.) are holdovers from outdated cognitive science frameworks and do not reflect how biological intelligence actually perceives
ML ignores integrative, ecological, and affective approaches to perception that are more likely routes to genuine intelligence and consciousness
These models are cheap workarounds that exclude the senses and represent homogenization posing as intelligence rather than true paths to vision understanding