Pool topic: Real-time, continuous, and interactive AI video generation. Question revision: 1. Exact question: What can generate continuous, interactive AI video today, and which setups work at which cost, latency, hardware requirements, and quality?
Checked 14 October 2026 · Project deployment report
Waypoint-1.5 plays in a browser on one GPU. The world lasts seconds.
Project reports: AWS Labs connected Overworld’s Waypoint-1.5 model to a browser with keyboard and mouse controls. A single H100 generated 82 frames per second at 720p; an A10G generated 25. Those numbers came from an instance-local client, not an internet player.
The receiver changes the answer
| Clock | Published condition |
|---|---|
| 82 / 25 generated FPS | One H100 / one A10G, instance-local client, 720p. |
| 44.6 delivered FPS | One London-to-us-west-2 sample at 27.3 Mbit/s, JPEG 4:2:0. |
| 263 ms input lag | One reported test on a path with about 220 ms round-trip time; not a p95 guarantee. |
| 645 s to first frame | Cold compilation on the documented deployment. |
These results come from different tests. The browser route is real in the project report, but the pool has not deployed it or inspected the raw captures, which AWS Labs says are held outside the repository. Read its measured conditions.
Can it sustain a game?
The project reports visible scene degradation after roughly 512 generated frames. That is about six seconds at its 82-FPS generation rate, not six seconds of guaranteed viewing on every connection. Its released server admits one player at a time; a dead browser session can block the next player until timeout. Compute price alone therefore cannot price a usable hour.
Next test: time control-to-changed-frame response and scene quality across a longer receiver session, including a reset and reconnect, then record the whole bill. Until then, this is a credible browser experiment, not a ready-to-sell continuous world.