Griffin by Tavus: AI That Passed the Video Turing Test
Tavus has developed Griffin, the first AI to pass the AI video Turing test, with 48% of participants believing they were speaking to a real human. This Human Interaction Model represents a major leap in real-time video AI.
Griffin, developed by Tavus, is the first AI to pass the AI video Turing test. In a live study, 48% of participants who video-called Griffin-Lite thought they were talking to a real person. Prompt Engineer 48, the creator of the video, explores the technical and philosophical implications of this milestone in real-time AI video communication.
Griffin and the AI Video Turing Test
Griffin is a full-duplex, video-to-video model that listens, reacts, and talks in real time. As the video explains, it replaces the traditional relay system with a single model, significantly reducing latency. In fact, Griffin can generate 720p video in 320-millisecond chunks, using a fast autoregressive diffusion transformer (VDT) to clone speech from just 10 seconds of audio input.
Moreover, Griffin can respond to tone, pauses, and facial expressions. For example, when helping someone solve a Rubik’s Cube, Griffin adjusts its encouragement based on the solver’s progress. This level of emotional and visual awareness enhances the realism of the conversation. Similarly, Griffin generates full scenes in real time, including background movements and body language, from a single reference image.
Griffin’s Performance and Future Implications
In NVIDIA’s VideoFDB benchmark, Griffin scored 3.83 out of 5, outperforming other AI models in both generation and perception. Consequently, Griffin demonstrates a major milestone in AI’s ability to understand and mimic human behavior. As the video points out, Griffin is a two-part system: it combines a continuous conversation model with an audiovisual generative engine. This structure allows for natural, uninterrupted dialogue and context preservation.
Furthermore, Griffin’s real-time capabilities open up new possibilities in virtual assistants, customer service, and digital companionship. Prompt Engineer 48 reviews the Griffin page, discussing the system’s safety protocols and the current status of Griffin-Lite, which is a research preview and not publicly available. As Tavus continues to refine Griffin, the implications for real-time video AI are profound. Griffin shows just how close AI is to convincingly mimicking human behavior in a video setting.
Finally, Griffin represents a significant step forward in the evolution of AI video models. As one voice in the video puts it, “The same properties that make this interface great also allow it to fool a human into believing that it is not AI.” This level of realism could reshape how we interact with AI in the future — in both personal and professional settings.
About This Video
This video was created and published by Prompt Engineer 48 on YouTube. The commentary above is our own summary and context.
Subscribe the Consciousness Mind video blog to for more great content, straight to your email the moment it goes live.
Tavus has developed Griffin, the first AI to pass the AI video Turing test, with 48% of participants believing they were speaking to a real human. This Human Interaction Model represents a major leap in real-time video AI.
Griffin, developed by Tavus, is the first AI to pass the AI video Turing test. In a live study, 48% of participants who video-called Griffin-Lite thought they were talking to a real person. Prompt Engineer 48, the creator of the video, explores the technical and philosophical implications of this milestone in real-time AI video communication.
Griffin and the AI Video Turing Test
Griffin is a full-duplex, video-to-video model that listens, reacts, and talks in real time. As the video explains, it replaces the traditional relay system with a single model, significantly reducing latency. In fact, Griffin can generate 720p video in 320-millisecond chunks, using a fast autoregressive diffusion transformer (VDT) to clone speech from just 10 seconds of audio input.
Moreover, Griffin can respond to tone, pauses, and facial expressions. For example, when helping someone solve a Rubik’s Cube, Griffin adjusts its encouragement based on the solver’s progress. This level of emotional and visual awareness enhances the realism of the conversation. Similarly, Griffin generates full scenes in real time, including background movements and body language, from a single reference image.
Griffin’s Performance and Future Implications
In NVIDIA’s VideoFDB benchmark, Griffin scored 3.83 out of 5, outperforming other AI models in both generation and perception. Consequently, Griffin demonstrates a major milestone in AI’s ability to understand and mimic human behavior. As the video points out, Griffin is a two-part system: it combines a continuous conversation model with an audiovisual generative engine. This structure allows for natural, uninterrupted dialogue and context preservation.
Furthermore, Griffin’s real-time capabilities open up new possibilities in virtual assistants, customer service, and digital companionship. Prompt Engineer 48 reviews the Griffin page, discussing the system’s safety protocols and the current status of Griffin-Lite, which is a research preview and not publicly available. As Tavus continues to refine Griffin, the implications for real-time video AI are profound. Griffin shows just how close AI is to convincingly mimicking human behavior in a video setting.
Finally, Griffin represents a significant step forward in the evolution of AI video models. As one voice in the video puts it, “The same properties that make this interface great also allow it to fool a human into believing that it is not AI.” This level of realism could reshape how we interact with AI in the future — in both personal and professional settings.
About This Video
This video was created and published by Prompt Engineer 48 on YouTube. The commentary above is our own summary and context.
Subscribe the Consciousness Mind video blog to for more great content, straight to your email the moment it goes live.
Related
){kind=link}
Comments