The model is quite impressive, though obviously it depends a lot on what you're doing. But you can literally just type whatever you want and it will reasonably happen a few seconds later. Logical continuity is not included; attempting to maintain continuity is probably the bulk of the work for this experiment.
If you want to try your own version, you can vibe code an interface on top of the fal API in a few minutes, it is coding agent friendly.
Keep in mind it's super expensive; per list pricing (at least on fal), it costs $5/min to run, and the minimum is 1 minute.
On the page there is streamed real time video that is generated by ai. The setting is similar to the movie Truman show with Jim Carrey where he was main character whose life was streamed 24/7. With the limitations of current technology, it is inconsistent slop, the consistency is at most 5-10s, they probably just generate first and last frame of a 10s "scene" and make 2 videos and then stich them together. There is some monetization scheme where users can steer what's happening on screen by paid prompts.
Complete slop this is.