EP 647 : Tavus Griffin AI Video Calls Fool 48% – Google Protein Watermarks, OpenAI & Microsoft
Imagine opening a video call and not being sure whether the person on the screen is human or an AI. In today’s episode we dive into Tavus’ groundbreaking Griffin model, which fooled 48 % of testers into believing they were speaking with a real person. We break down how Griffin‑Lite watches, listens, and interjects mid‑sentence, its near‑human score on NVIDIA’s VideoFDB benchmark, and the safety safeguards Tavus is building before a wider rollout.We also explore Google DeepMind’s new SynthID Bio, a proof‑of‑concept that embeds invisible watermarks into AI‑generated protein structures…
The skinny
The skinny isn't ready yet — notes appear once the transcript is processed.