Dorsey's Block Debuts Human Conversation AI Model Dessa Trained From Scratch

Jack Dorsey’s Block has introduced Dessa, an AI model trained from scratch on human conversation alone, positioning it as a speech-to-speech system built without text-based training data. To demonstrate its capabilities, the company shared several sample exchanges that highlight both natural dialogue flow and occasional awkward, uncanny moments. In one interaction, the model mistakenly calls the speaker “Ronaldo,” while another clip shows it taking part in a voice-based guessing game. The Dessa website includes additional demonstrations, including a conversation about Christopher Nolan’s films that drifts into an unexpected discussion of Bryan Singer’s Superman movies, as well as a reply to a user planning a surprise trip to Paris that comes across as emotionally expressive. The site also compares Dessa with OpenAI’s GPT-Live-1 voice model, suggesting that Dessa sounds more natural in some exchanges. Block describes Dessa as a true speech-to-speech model whose knowledge comes entirely from human speech. The company says it can run locally on a MacBook Pro M4 Max or better, signaling an emphasis on accessibility for personal computing environments. Block said the model was created for three main reasons: curiosity, improving understanding between humans and AI, and advancing AI safety. It added that a small group of AI safety researchers will receive early access to the model.