Block made an AI model that talks like an awkward friend, and Jack Dorsey has a safety goal

A model trained on conversation
Block says Dessa was trained “from scratch on nothing but human conversation.” The company shared recordings of people talking with the model to demonstrate how it responds. Dessa lead Ragavan Thurairatnam shared one exchange that has a mostly natural back-and-forth. At first, however, the model mistakes the speaker for someone named “Ronaldo.”
Chats take unexpected turns
Other examples on the Dessa website include a discussion of Christopher Nolan’s films that veers toward Bryan Singer’s Superman films. In another, a person describes a surprise trip to Paris. The website says Dessa responds to the Paris story with “what sounds like teary gratitude.” A separate conversation with OpenAI’s GPT-Live-1 voice model makes Dessa seem more natural, according to the article.
Speech, emotion and safety
The team lists curiosity, better understanding between humans and AI, and safety as reasons for building it. It says voice AI may miss emotion in a person’s voice, including signs of serious pain.
Early access for researchers
The team says babies learn to speak without seeing written words and wanted to test whether a model could do the same. It says that is possible, but very hard.
Block plans to provide early access to a small group of AI safety researchers. The team describes Dessa as a step toward addressing voice AI’s difficulty recognizing emotion.
Key points
- Block says Dessa was trained from scratch using human conversation.
- Shared examples include natural exchanges as well as awkward or unexpected turns.
- The team says the model can run locally on a MacBook Pro M4 Max or better.
- A small group of AI safety researchers will receive early access.