Jack Dorsey’s Block has built a voice model called Dessa. The company says it trained the system from scratch on nothing but human conversation. No text corpus, no scraped web pages, no instruction dataset. Just people talking to each other.
Block published sample conversations to show what that produces. The results are uneven in a way that feels familiar. Dessa keeps a mostly natural rhythm, then trips over a detail a person would never miss. In one clip, the model mistakes the speaker for somebody named Ronaldo.
Ragavan Thurairatnam, who leads the Dessa effort, shared that exchange publicly. A second demo shows the model playing a guessing game by voice. It hesitates, guesses wrong, and recovers. The pauses land close to where a human pause would land.
The Uncanny Parts Are the Point
The Dessa website collects more examples. The model discusses Christopher Nolan’s films, then drifts into an unexpected tangent about Bryan Singer’s Superman movies. In another clip, a person describes a surprise trip to Paris. Dessa answers with what the site calls “teary gratitude.”
Nobody scripts those moments. They emerge from training on conversation rather than on written text. Speech carries hesitation, interruption, and repair. A model fed only transcripts of talk learns those patterns directly instead of inferring them.
Block also posted a conversation between Dessa and OpenAI’s GPT-Live-1 voice model. Listeners described Dessa as the more natural of the two. That comparison is the real claim here. Block is arguing that conversation data beats raw scale for voice quality.
Why a Payments Company Is Doing This
Block runs Square and Cash App, businesses built on phone calls and customer support. A model that handles spoken conversation well has obvious commercial use. It also gives Dorsey a research story separate from the payments business.
The demos are curated, and curated demos always flatter the company that made them. Nobody has published independent benchmarks for Dessa. Block has not said when the model ships or who gets access. For now the company is showing personality rather than performance numbers.
That may be enough to start a wider argument about how voice AI should be built. Training on human speech instead of text is a genuine fork in the road. Dessa’s awkwardness is evidence that the path produces something different from the usual assistant voice.
Source: The Verge — Jack Dorsey’s Block made an AI model that talks to you like an awkward new friend
Related reading: OpenAI Now Trains Its Frontier Models With No Internet Access and Canva Launches Craft, Its First In-House AI Image Model.
The Netherlands, for internationals
One email a week: the week's news for internationals in the Netherlands, new practical guides and what changed in the cities. No spam, unsubscribe any time.
