# From tweet to conversation: this is Dessa, Jack Dorsey's AI that listens while you speak

> Dorsey unveils Dessa, an AI trained only on human conversations that listens and speaks at the same time and runs on a MacBook Pro M4 Max
> By Covadonga Fernandez · Artificial Intelligence · 2026-10-24

Jack Dorsey could once again change how we communicate with Dessa, an AI that learns to speak by listening to human conversations and allows users to talk to it by voice, with natural pauses, nuances and interruptions, according to its developers.

Block, the company Dorsey runs, presented yesterday, October 23, Dessa, an artificial intelligence model trained from scratch exclusively on human conversations, without text or previously trained models. According to its developers, it can listen and speak at the same time, respond with natural intonation and pauses, and run on a MacBook Pro M4 Max or higher.

Dorsey announced the project [on his X account](https://x.com/jack/status/2113697902916751661), where he highlighted both the naturalness of the conversations and the possibility of using it locally. The team's proposal is that interaction with an AI should preserve information that a transcript does not capture. For example, how the meaning of a sentence changes depending on tone, breathing, a laugh or a silence.

In the official presentation of Dessa, signed by the head of the lab, Ragavan Thurairatnam, the researchers explain that they wanted to test whether a machine could learn to converse by listening, as children do before they can read. They claim to have achieved this through training based solely on human speech.

## An AI that listens while it speaks

Dessa receives audio and produces audio. It is a model of simultaneous two-way communication, which allows it to listen to the interlocutor while generating its own voice. This capability aims to overcome one of the difficulties of conversations with assistants, namely having to wait for one party to finish before the other responds. According to the team, Dessa has learned when to interject and when to accompany the speaker with brief expressions indicating that it is following the conversation.

Block maintains that the model also processes signals such as sarcasm, sighs of frustration or a voice breaking with sadness. Its goal is to preserve those nuances instead of reducing the interaction to transcribed words.

The company presents this capability as both an improvement in the experience and a safety matter. The example it uses is that of a person expressing pain while an assistant keeps talking without acknowledging it. Dan Garon, investment partner at The Residency and former Zynga executive, believes this technology anticipates how small businesses will be able to converse with their customers and handle their tools by voice. On his X account he notes that Dessa is part of a Block bet on connecting AI assistants, automated payments and models that run on the devices themselves.

## Conversations with laughter, pauses and interruptions

The website includes recordings that, according to Block, correspond to live conversations without a person directing the system's responses. Dessa decides what to say, when to interject and how to express itself. In one of the demonstrations, the model laughs during a sentence from the interlocutor. In another, it helps complete a recollection of a film. Changes in intonation, responses to a surprise and a conversation with another artificial intelligence are also shown.

[Microsoft's AI doubles doctors' success rate in complex diagnoses](https://observatorioblockchain.com/en/ai/ia-microsoft-duplica-aciertos-medicos-casos-complejos/)

Dorsey stresses that Dessa is small enough to run on a laptop. The presentation specifies that it must be a MacBook Pro with an M4 Max chip or higher. The announced capability therefore corresponds to high-performance equipment.

The researchers also claim that Dessa learned language with hundreds of times fewer words than large language models, although they acknowledge that the latter are also trained for different tasks, such as programming. The presentation does not detail a comparison that would allow the general capabilities of both types of systems to be equated.

The scientific interest of the project lies in exploring how much an AI can learn directly from speech, from language and knowledge to conversational turns. The team also plans to investigate how it could acquire new languages and learn from the information it receives during an interaction.

For now, the company offers early access to a small group of security researchers to identify risks the team has not anticipated.
