Tavus Launches Griffin, First Model to Pass Video Turing Test
- On Oct. 1, Tavus introduced Griffin, a Human Interaction Model capable of real-time video and audio processing.
- In a test with 54 participants, nearly half (48%) believed Griffin-Lite was a real person after a one-minute video call.
- Griffin-Lite achieved the highest ranking in NVIDIA VideoFDB testing among 15 models for full-duplex audiovisual interaction.
- The model recorded an impressive average latency of just 0.43 seconds on H100 GPUs during performance tests.
- Access to Griffin is currently limited due to concerns over its ability to impersonate humans and associated security risks.
Griffin represents a significant advancement in AI technology, showcasing capabilities that allow it to interact seamlessly with human users while responding to nonverbal cues and context. Tavus is focusing on developing AI disclosure features and assessing risks before wider public release.
With nearly half of the participants mistaking Griffin-Lite for a human, this model sets a new benchmark in AI-human interaction technology.(Source)
