Block presented Dessa on October 23, 2026, a voice model developed by its AI research lab led by Ragavan Thurairatnam. Trained exclusively on human conversations, with no books, web pages or text, the model operates end to end and in full-duplex mode, allowing it to handle interruptions, laughter, pauses and the emotional rhythm of exchanges. The demonstrations impressed observers with their naturalness but also revealed awkward errors, such as calling a participant “Ronaldo.” The model is not broadly available, and Block is prioritizing AI safety researchers with early access to study impersonation, trust and other risks before a wider launch. Dessa also extends Block’s work on the open-source “goose” agent into the conversational layer of AI.
Source: Read the original article
