Meta's live staged demo fails; the "AI" recording plays before the actor acts (old.reddit.com)

Meta’s staged “live” demo of an AI-driven actor stumbled when footage intended to be generated in real time visibly played ahead of the human performer, revealing the clip was pre-recorded or that the playback pipeline had severe synchronization problems. Viewers noticed the system’s audio/visual output occurring before the actor moved, a telling sign that the demo either relied on canned material or that the supposed real-time inference and rendering stack was compensating with lookahead buffering or mis-timed feeds. This matters for AI/ML because demos are a primary way researchers and companies communicate model capabilities; misrepresentations—intentional or accidental—undermine trust and distort expectations about what current generative-video systems can do. Technically, the failure points to differences between pre-rendered outputs and true low-latency inference: issues could stem from precomputed video assets, pipeline caching, latency compensation algorithms, or faulty synchronization between capture (motion, audio) and synthesis modules. The incident reinforces calls for transparent, reproducible demos, clear distinction between pre-recorded and live outputs, and standardized benchmarks for latency, fidelity, and robustness when claiming real-time generative capabilities.