Infinite Slop launches live chat-controlled AI video stream

• Chat prompts now act as a live programming queue: in levelsio’s build post, each generated segment is connected to the clip before it.
• The stream can generate 15 seconds of video in nine seconds, according to levelsio’s speed report, leaving headroom for a perpetual broadcast.
• Viewer submissions are queued and throttled: levelsio’s queue update added progress visibility, while a reply from levelsio caps overloaded periods at four 15-second clips per minute.
• The experiment blocks NSFW and copyright requests, levelsio’s moderation reply says.
• Continuous generation carries a serious compute bill: levelsio’s estimate puts the run at roughly $4,000 per day, sponsored by fal.

fal’s H3 Max launch post says the model and inference stack were co-designed and trained and served on NVIDIA GB200 NVL72 systems. The live channel exposes three production states, NOW GENERATING, UP NEXT, and ON AIR.

Levelsio’s loop is simple: a chat message becomes the next scene, and the generator uses the preceding video to connect it to the new one.

That continuity layer also carries artifacts forward. Levelsio said recurring elements can appear because the next job uses the previous video in a reply. When viewers repeatedly entered “banana,” the bridge between prompts made banana imagery take over the broadcast.

Levelsio’s claim, 15 seconds of output in nine seconds, maps to three five-second renders at about three seconds apiece.

Fal’s launch post reports roughly 35 times the throughput of the official MiniMax H3 endpoint for a five-second clip. The post also quotes Design Arena’s benchmark as finding more than 50 times the speed, while fal attributes the model gains to post-training with additional data focused on prompt adherence and visual quality.

Quality remains a separate judgment. Ozansihay said H3 Max keeps the same quality and synthetic look as base H3, while arguing that Seedance still leads visually in ozansihay’s comparison.

Creator examples in the evidence use H3 Max inside wider generation stacks:

• In gokayfem’s post, a single prompt through fal Agent produced a clip combining H3 Max and Nano Banana Pro in two minutes.
• Gokayfem combined Krea 2 Turbo, ElevenLabs music, and H3 Max; in gokayfem’s workflow, video stitching and audio merging had become the bottleneck.
• A heyglif skill by fabianstelzer generates motion-graphics explainers from a script, idea, or prompt.
• Ozansihay tested three modes with Turkish-dialogue scenes: Start-End Frame, image-to-video, and text-to-video in ozansihay’s mode test.

During busy periods, levelsio’s reply says the system selects only four prompts each minute, exactly enough to fill a minute with four 15-second clips.

Levelsio told one commenter that the prompt behind each generated video appears in the top-left corner. The creator also floated a Neuralink-connected version in a brief reply.

The current controls include content restrictions. Levelsio said NSFW or copyrighted prompts are blocked in a moderation reply, and separately said copyrighted material is not allowed in another reply.

In levelsio’s cost estimate, the creator put the stream’s run rate at about $4,000 a day and said fal was sponsoring it.

A levelsio stream update reported about 400 people watching during an early run.