It's 2027, and imagine you record a 60 sec video of a dog in the park. Now imagine you take the first frame of that video and give it to the most powerful video model ever created at that time and ask it to create a 60 sec video only with that frame. If the current pace of video models continues on the trajectory we're seeing, you can expect that by then (or in a not-too-distant future), you will, theoretically, be able to generate a video that's identical to the original 60 sec video. The generated video is equivalent to the real video. The motion, physics, movements, actions, and all possible outcomes of every possible scenario of that video are somewhere in the latents of the model. If you can find the one that represents the reality you experience then you have a generated video that's the same video you recorded. But you can also, theoretically, create every variation of your original video with any possible permutation possible. That means we have finally created a simulation of reality. The map is the territory. Now imagine it's 2028 and you can do all of this in real-time...