Shortly after DeepMind released the Veo 3 paper, some were quick to declare it the GPT-3 moment for visual reasoning. The paper claims that through massive video training, Veo 3 demonstrates zero-shot learning capabilities, enabling it to solve tasks not explicitly included in the training data.


However, a closer look reveals several issues. First, this is not the original Veo 3 paper but an analysis report on its capabilities. Second, the so-called "Chain-of-Frames" concept is essentially a rebranding of the language model's "Chain of Thought" applied to video models. The comparison image is telling—the same person's two expressions, paired with old and new terminology, are packaged as a breakthrough.


Tesla engineers put it bluntly: the real test lies in handling messy real-world videos, not clean lab data. How does Veo 3 perform in continuous 360-degree object rotation tests? The paper doesn't mention it. Meta's UniBench research points out that such visual reasoning capabilities are not stable.


The most interesting observation comes from the comments: while everyone is busy discussing "emergent" properties, no one is asking whether these capabilities truly go beyond pattern matching. Veo 3 might pass certain tests, but it's still far from genuine physical understanding and causal reasoning.

Autonomous driving requires stable and reliable visual reasoning, not pretty curves in the lab. DeepMind's fireworks this time are dazzling, but to illuminate the complexities of the real world, sparks alone aren't enough.