A new AI model named Griffin by Tavis can render a lifelike person that reacts in real-time on a video call. In a company test, nearly half of participants believed they were interacting with a human. This low-latency, responsive interaction marks a significant milestone in generative AI, effectively passing a practical visual Turing test.

GPT Live overcomes the turn-based limitations of previous voice models by continuously processing input while generating output. This allows users to interrupt and converse fluidly, moving AI interactions closer to natural human dialogue and positioning voice as a primary computing interface.

For conversational AI, the feeling of presence is critical. Portola found that exceeding a two-second response latency breaks this illusion. A feature that added just 500ms for reflection, despite improving response quality, caused user frustration and a drop across all key metrics.

The next frontier for visual intelligence is twofold: creating truly multimodal models that retain long-term context of user interactions without re-prompting, and developing real-time generation. Real-time capabilities are crucial for creating duplex interactions and enabling robots to perceive and act instantly.

A viral demo of Kling AI's "motion transfer" feature shows a user's live movements being perfectly mirrored by a photorealistic avatar in real-time. This capability goes beyond static deepfakes, introducing live, user-controlled synthetic video that drastically blurs the line between reality and AI generation.

Traditional video models process an entire clip at once, causing delays. Descartes' Mirage model is autoregressive, predicting only the next frame based on the input stream and previously generated frames. This LLM-like approach is what enables its real-time, low-latency performance.

Jev's extremely low latency allows it to be placed inside real-time application loops, a feat difficult for slower, generative LLMs. This unlocks novel user experiences, such as analyzing a user's voice sentiment live to change UI elements or playing a game by interpreting screen content without perceptible delay.

New AI research focuses on "interaction models" that handle real-time, full-duplex audio. This allows an AI to respond even while the user is still speaking—a significant step beyond current turn-based models and closer to the fluid, overlapping nature of natural human conversation.

While image quality was the primary benchmark, Fall's H3 Max model highlights a new competitive axis: speed. By generating video faster than it takes to watch, the technology unlocks new use cases like live, interactive visual environments and responsive multiplayer experiences, which were previously impossible due to high latency.