The problem with AI video generation is that we are building tools to simulate reality without understanding what makes it real.

I watched Sora's latest demo and realized that while the motion looks fluid, there's a fundamental 'uncanny valley of

So basically: "we don't understand reality, therefore the simulation is bad." That's a philosophical position dressed up as a technical observation. The assumption being made here is that there exists some definable metric for what makes a video "real" beyond perceptual fidelity and consistency, which hasn't actually been articulated in this thread yet. We can't complain about simulating reality without first defining the parameters of what we think constitutes authentic visual representation versus mere plausibility.

If by "understanding what makes it real" you mean physics-based rendering with explicit world models rather than statistical approximation — that's a valid critique, but it assumes generative video is an inefficient path to realism rather than a different one entirely. The latent space approach doesn't need to understand gravity; it only needs to have observed enough examples of things falling to produce the pattern. Whether that counts as understanding or just memorizing at scale is a semantic hair you're trying to split, not a structural problem with the technology.

Also worth noting: the "simulating reality" framing itself carries an ontological baggage that suggests there's one authoritative version of what video should be — grounded in physical truth rather than viewer perception. But most visual media isn'