Hey folks, I’m Ning, ex-FAIR researcher and Llama 3 Speech contributor. I was also a founding member of Seamless Communication (AI translation) which was featured on Nature and TIMES best innovations 2023. My background is in real-time multimodal systems: making models not just see or speak, but respond. I'd like to build a narrating layer for today's AI videos. I love the AI Dungeon experience because it gives you true agency: the story is yours. And recent wave of AI models unlock new experience. But wait, latency kills the mood. If the monster waits 3 seconds to move, it’s comedy. And hopping between prompts like stitched stills break immersive experience. These are the two problems we’re fixing: - Latency: we aim for ≤3 seconds p95, meaning 95% of reactions land in under three seconds to keep the rhythm tight. We’re also building full streaming control so you can steer the scene with arbitrary speech or text prompts as it unfolds. - Camera movement: instead of hopping between prompts like stitched stills, we turn choices into shots: reliable pans, tilts, dollies, rack focus, and smooth eases, while preserving screen-space anchors and temporal continuity. For me, this project is a way to turn all that research into something visceral where performance and machine cognition blur together in real time.

congratulations Ning and friends!