Developed by the AI research lab Tavus, it represents a shift from text- or voice-only chatbots to an integrated video-to-video AI experience. Griffin can see, hear, interpret, and generate natural responses in real time. Instead of operating like a standard voice assistant with a digital face slapped on top, it dynamically reads the user’s hand movements, conversational timing, and physical objects held up to the camera. Key Capabilities & Technical Features • Full-Duplex Interaction: Unlike older AI systems that require users to wait for a turn to speak, Griffin processes incoming video and audio concurrently. It understands conversational cues, nods, laughs, handles interruptions smoothly, and can even interrupt the user naturally. • Complete Scene Generation: It renders full video frames in real time, from realistic facial expressions and hand gestures to lighting adjustments and subtle movements like swiveling in an office chair. • Top Benchmarked Performance: It currently ranks #1 on NVIDIA’s Video Full-Duplex Benchmark, outperforming competing video AI systems by roughly 37%. The "Video Turing Test" Results Tavus introduced the model alongside a blind study where 54 participants had a brief, one-minute video call with an AI avatar running on a version called Griffin-Lite. Nearly half of the participants—48%—believed they were interacting with a real human being. By comparison, Tavus's previous generation AI system only fooled 2.4% of people under the same testing parameters. Availability & Use Cases Tavus designs the underlying models that power customizable digital representations known as "PALs" (human avatars used for specific tasks). Potential commercial use cases include live AI tutors who can physically sense when a student is struggling, realistic role-play tools for employee training, or customer support agents who can visually identify broken components on camera. Because the technology is convincing enough to raise concerns regarding deepfakes and impersonation, Griffin is not yet publicly available. It is currently limited to trusted research preview testers while Tavus develops specialized safety features and disclosure frameworks. https://finance.yahoo.com/technology/ai/articles/tavus-introduces-griffin-first-face-163000466.html