Next crazy thing: accurate time evolution of photographed phenomena
"DALL-E + 0.001ms" ran continuously = reality sim
Based on previous frames, it'd be able to predict that two water drops would coalesce within the next dozen frames, a car will stop within X distance of another, etc.
oh yeah, this will be crazy. DALL-E basically does text-to-image; there's a whole area of text-to-video that's working on exactly what you're talking about.
There are some good examples from a recent paper here: https://video-diffusion.github.io/ they generate timelapses of fireworks, rivers, pouring liquids, etc.
"DALL-E + 0.001ms" ran continuously = reality sim
Based on previous frames, it'd be able to predict that two water drops would coalesce within the next dozen frames, a car will stop within X distance of another, etc.
Next frame probability Scheherazade