What Dessa is and how it was trained
Block, the company led by Jack Dorsey, has presented an artificial intelligence model called Dessa, trained “from scratch on nothing but human conversation.” The news is reported by The Verge, which cites the project team’s website and the materials it released. According to Block’s description, Dessa is a “true speech-to-speech” model—that is, designed to go directly from speech to speech—and everything it knows derives from human speech. According to the information provided, the model can be run locally on a MacBook Pro M4 Max “or higher.”
The project team cited three reasons behind the work: curiosity, meaning the fun of trying; better understanding between humans and artificial intelligence; and safety. The material released discusses the idea of verifying whether a model can learn to speak without ever seeing a written word, as happens with a child, adding that this is possible but “very difficult.”
The demonstrations and the limitations that emerged
To show how the model works, Block shared some conversations with humans, described at times as embarrassing or unsettling. In one example shared by Ragavan Thurairatnam, the head of Dessa, the model initially mistakes the interlocutor for a person named “Ronaldo,” although it subsequently maintains a largely natural back-and-forth. A video shows Dessa taking part in a spoken riddle game.
The project website collects other examples: a conversation about Christopher Nolan’s films, with a digression into the Superman films directed by Bryan Singer; a person telling Dessa that they brought it to Paris as a surprise, to which the model responds, according to the website’s description, with what appears to be moved gratitude; and a dialogue between Dessa and OpenAI’s GPT-Live-1 voice model, in which Dessa appears more natural. Some excerpts from these conversations are available in a video.
The examples serve to illustrate the model’s potential, but they also show that its behavior is not yet entirely consistent: the name error and the digressions are described in the same materials released by Block and should be read as selected demonstrations, not as a systematic evaluation of performance. The material does not indicate metrics, model size, or comparative test results.
The reasons for the project and next steps
According to the team, the safety motivation concerns the difficulty current voice systems have in accounting for emotion in the voice. The material gives the example of a person experiencing serious pain while a voice system does not notice and keeps talking: this, according to the team, is the safety problem Dessa aims to help solve. It is also argued that speech is the most natural interface for any intelligence and that current voice AI wastes it, flattening speech into text and failing to keep pace with the timing of human conversation.
Block announced that it will give early access to the model to “a small group of AI safety researchers.” No timeline, selection criteria, or access arrangements are indicated. It therefore remains to be seen how the model behaves outside the published demonstrations, what its performance is on emotion and safety, and how access for researchers will be managed.
The question of comparison with existing voice models
The only comparison in the material is the dialogue between Dessa and OpenAI’s GPT-Live-1, in which Dessa is described as more natural. This is, however, an example chosen and published by Block, not an independent test: the comparison should therefore be attributed to the company and not treated as a third-party evaluation of the two models’ capabilities.
The relevant point, in light of the available facts, is that Dessa proposes a different approach: a model trained only on speech and designed to run locally, rather than converting voice into text. If access for safety researchers produces public evaluations, it will be possible to understand whether this approach translates into measurable advantages. For now, Block’s announcement makes known a prototype presented through curated examples and team statements, while independent evidence remains to be produced.
