Research update 007 · 16 September 2026
YoLive has a real-time claim. The useful question is who gets to choose the next scene.
Yoroll says its new YoLive product turns audience proposals and votes into the next generated scene. Its H3 Superfast claim is 10 seconds of 768p, 24 fps video with native audio in four seconds on eight B200 GPUs. That clears a clip-generation threshold. It does not yet tell us how long a voter waits to see the result.
This is a public group-choice route, not a generic video model launch
Provider says: viewers of YoLive propose what should happen, vote on the direction, and the winning proposal guides the next scene. Yoroll presents H3 Superfast and the product together, with an initial Zombie Scavenger channel and other planned interactive formats.
That makes the route useful to study. The stated interaction is collective and scene-to-scene: a group chooses the next beat. It is not evidence that one person can alter the current frame or freely steer a held session.
Four seconds is a generation figure, not a voter’s wait
Provider says: H3 Superfast generates 10 seconds of 768p, 24 fps video with native audio in four seconds on eight NVIDIA B200 GPUs. By simple duration arithmetic, that is 2.5 times playback rate and leaves six seconds of gross lead while a 10-second clip plays.
A usable group loop must spend that lead on collecting proposals, closing the vote, moderation, turning the winner into an instruction, scheduling, delivery, and retries. The launch announcement does not publish those timings, a tail-latency record, concurrent capacity, cost, or public uptime.
Compare the control contract before comparing speed
| Route | Who changes the story? | What the source supports | Missing proof |
|---|---|---|---|
| YoLive | A group proposes and votes on the next scene. | Provider says: 10 seconds in four seconds on eight B200 GPUs. | Vote-to-screen delay, moderation behavior, scene persistence, costs, and observed public use. |
| H3 Max Director | A holder of an open session sends directions while the stream plays. | Provider says: a WebRTC session carries context across segments. | Measured direction-to-visible-change time, handoff, uptime, and total delivery cost. |
| One-5090 FastH3 channel | The operator prequeues independent scenes. | Community project says: 22.1 fps at 448×448 with retimed playout. | Viewer control, native 24-fps motion, story state, replication, and long-run economics. |
The next benchmark needs timestamps, not another clip clock
- When did the proposal window open and close?
- Which selection rule and moderation step chose the instruction?
- When did that instruction reach the generator?
- When did the first visible change from it reach the viewer?
- Did the new scene preserve the world, character, and winning choice after a retry or restart?
Until that record exists, call YoLive what the source supports: a provider-announced collective next-scene product with a source-stated faster-than-playback generation condition on eight B200 GPUs. It is a promising deployment shape, not a measured proof of instant or durable interaction.