The AI research team at Tavus has introduced a groundbreaking real-time video interaction model called Griffin. During a trial run, its preliminary version, Griffin-Lite, was nearly indistinguishable from a real human by almost half of the participants in a one-minute video call experiment. This achievement marks a significant leap over previous systems. Tavus proudly proclaims that Griffin is the inaugural model to successfully pass the real-time video Turing test, positioning it as a pioneering model for human-like interaction.

The model adopts a full-duplex video interaction approach, revolutionizing the interaction logic that was previously limited to traditional cascaded systems. It has the capability to simultaneously perceive non-verbal cues, including users' facial expressions, vocal tones, and pauses. Moreover, it can generate complete 720p videos in real-time, complete with character movements and environmental details, all while maintaining exceptional latency performance.

In a series of public benchmark tests, Griffin's performance was on par with human levels, outshining other similar AI systems. Nevertheless, it's important to note that the experiment had its limitations, such as a relatively small sample size and a brief call duration. Consequently, the conclusions drawn from this experiment necessitate further validation.

At present, Griffin is exclusively available for preview to a carefully selected group of trusted testers. The Tavus team is diligently working on developing an AI identity disclosure feature and advancing safety alignment evaluations. As real-time digital human interactions increasingly mirror those with real people, it becomes paramount to clearly inform users about the identity of their interaction partners. This transparency will be crucial in fostering trust and confidence in the usage of such technologies.