Jack Dorsey’s Block made an AI model that talks to you like an awkward new friend

If you use voice assistants, their ability to understand emotional tone and natural speech timing is likely to improve significantly.

Jack Dorsey's company Block has developed an AI model named Dessa, which is trained exclusively on human conversations and operates as a speech-to-speech system. Demonstrations showcase Dessa engaging in somewhat awkward but generally natural-sounding dialogue, sometimes mistaking identities or going on tangents. The model can run locally on high-end MacBooks, with its training rooted entirely in auditory human interaction, mimicking how babies learn to speak. The development team cited curiosity, improved human-AI understanding, and AI safety as motivations. Dessa is designed to interpret and respond with a natural rhythm and to detect emotional cues in human speech, addressing a perceived limitation in current voice AI. Early access will be provided to a select group of AI safety researchers.

Dessa's development signifies a move towards more naturalistic and emotionally resonant AI interactions, potentially redefining user interfaces. By training solely on spoken language, Block aims to overcome the limitations of text-based AI, enabling machines to grasp nuances like timing and emotion that are critical for genuine communication. This focus on speech as the primary interface challenges the current paradigm where voice is often converted to text, losing vital contextual information.

The emphasis on AI safety, particularly the model's ability to detect distress, suggests a growing industry concern for more empathetic and responsive AI systems. Providing early access to AI safety researchers indicates a potential strategy to address ethical considerations and build trust around these advanced conversational AIs. This approach could lead to voice assistants that are not only more capable but also more attuned to user well-being.