Jack Dorsey’s Block has introduced Dessa, an AI model it describes as trained “from scratch on nothing but human conversation.” No books, no web pages, no text of any kind. Just people talking.
The pitch is realism. The demos show a model that handles timing, interruptions, laughter and pauses with surprising ease. They also show it calling a user “Ronaldo” for no clear reason, which is a very human mistake in its own way.
What Block actually built
Dessa launched on October 23, 2026, through a page at block.xyz/dessa and a round of posts on X. Block’s internal AI research lab is behind it, led by Ragavan Thurairatnam.
The name has history. Thurairatnam co-founded the original Dessa, an AI startup that Block acquired in 2020. Several members of that team now staff the lab, so the brand has come full circle as a product name.
The model is speech-to-speech. Audio goes in and audio comes out, with no detour through written words along the way.
Its defining trick is that it runs in full-duplex mode. Picture the gap between a walkie-talkie and a phone call. With a walkie-talkie, one person talks while the other waits for “over.” On a phone call, both sides can speak and listen at once, which is how real conversation works.
Dessa aims for the phone call. It can listen while it speaks, so it can react when someone cuts in, trails off, or laughs at the wrong moment. According to Block’s materials, the model is built to manage emotional tone and the rhythm of a back-and-forth, not just the words.
The no-text decision is the headline technical bet. Block says it deliberately left out text-based pretraining to better understand what it calls speech-native intelligence. The idea is a model that learns language the way a toddler does, by hearing it, rather than reading a library first and then learning to pronounce it.
Natural, mostly
Block chose to show its work. Thurairatnam shared sample conversations between Dessa and real people, and the results land somewhere between impressive and slightly eerie.
In one widely circulated exchange, the conversation flows with a mostly natural back-and-forth. Then the model opens by mistaking the speaker for someone named “Ronaldo.” The user, presumably, was not Ronaldo.
Early reactions have split along familiar lines. Some observers have described Dessa as a potential “GPT-3 moment” for speech, a nod to the release that made many people take text-generating AI seriously. Others pointed to stretches of awkwardness in the demos that pull the experience back into uncanny territory.
The demos also covered lighter scenarios, including interactive games. That choice makes sense for a showcase. Games test timing and turn-taking in a way that a scripted question-and-answer session does not.
Dorsey weighed in on X, saying he was proud of the team and the progress it had made.
A cautious rollout
Dessa is not broadly available. It is in early access, and Block is prioritizing AI safety researchers in that first group.
The stated goal is to study the model’s implications and risks before a wider release. That is a notable choice for a voice model specifically. A system that sounds convincingly human, reacts to emotion and handles interruptions raises obvious questions about impersonation and trust that a text chatbot does not.
Engaging the safety community up front also fits the novelty of the approach. Training only on speech is a different paradigm, and Block appears to want outside eyes on how a model like that behaves before it shows up in the wild.
Why Block is doing this at all
Block is better known for payments products than for AI research labs. But Dessa is not its first move in the space.
The company has previously worked on an open-source AI agent codenamed “goose.” Dessa extends that investment into a different layer of the stack: not what an AI does on your behalf, but how you talk to it.
Disclosure: This article was edited by Diego Almada Lopez. For more information on how we create and review content, see our Editorial Policy.