Editorial illustration for Block's AI model Dessa, trained from scratch, talks like an awkward new friend
Block's Dessa AI Model Trained on Human Talk
Block's AI model Dessa, trained from scratch, talks like an awkward new friend
Block put out an AI voice model called Dessa this week, and the pitch is unusual for the industry: it's trained from scratch on nothing but human conversation, not text scraped from the internet. Ragavan Thurairatnam, who leads the Dessa project at Jack Dorsey's company, has been sharing sample conversations to show what that approach sounds like in practice. The results land somewhere between impressive and unsettling.
In one clip, the model and a human speaker fall into a back-and-forth that feels mostly natural, right up until Dessa inexplicably addresses the person as "Ronaldo." Other demos on the Dessa website include a rambling chat about Christopher Nolan's films that veers into Bryan Singer's Superman movies, and a conversation where someone tells Dessa about a surprise trip to Paris. Block describes the model's reaction to that one as something close to teary gratitude. There's also a recorded exchange between Dessa and OpenAI's GPT-Live-1, pitting the two voice models against each other directly.
Block calls Dessa a true speech-to-speech system built entirely from spoken language, with a design meant to run locally rather than through the cloud.
Jack Dorsey’s Block announced an AI model called Dessa that’s trained “from scratch on nothing but human conversation,” and to show off the model, Block shared chats it has had with humans that can occasionally be a little awkward and uncanny.
Why this matters
Block isn't the first company to promise a more "human" chatbot, but training Dessa from scratch on conversation alone, rather than fine-tuning an existing foundation model, is a different bet. Most labs lean on internet-scale text and then bolt on alignment tricks to make outputs feel personable. Dorsey's team is testing whether conversation itself, stripped of the usual web-scraped noise, produces something that talks more like a person and less like a search engine with manners.
The awkward, uncanny demo clips Block is sharing aren't a bug they're hiding, they're the pitch. For developers and founders, that's worth watching closely: if a narrower training diet can produce a model with distinct conversational texture, it suggests personality and tone might be engineerable through data selection rather than endless parameter scaling. For researchers, the open question is whether "awkward" readings are a sign of genuine novelty or just undertrained weirdness.
Either way, a payments company building its own model from the ground up, instead of licensing one, says something about where Block thinks differentiation in AI products is actually going to come from.
Related Reading
Business & Startups Nordic pilot adds Gemini for Education, NotebookLM to boost AI literacy
Business & Startups Kling launches Video O1, all-in-one model with MVL bridge using transformer
Business & Startups DeepSeek Seeks More Capital Weeks After USD 7B Funding Round
From Archives LTM Launches BlueVerse Industrial AI Lab in Chennai to Push Pilots to Production
From Archives Infosys AI revenue rises 29% in Q2; FY27 guidance cut; USD 5.2B TCV
Common Questions Answered
What is unique about Block's Dessa AI model compared to other AI voice models?
Dessa is trained from scratch exclusively on human conversation data rather than internet-scraped text like most other AI models. This training approach represents a different bet from typical labs that use internet-scale text and then apply alignment techniques to make outputs feel more personable.
Who leads the Dessa project at Block and how is the model being demonstrated?
Ragavan Thurairatnam leads the Dessa project at Jack Dorsey's Block company. He has been sharing sample conversations between Dessa and human speakers to showcase what the conversation-only training approach produces in practice.
Why does Block's approach of training Dessa on conversation alone differ from most other AI labs?
Most AI labs rely on fine-tuning existing foundation models trained on internet-scale text and then add alignment tricks to make outputs feel more human-like. Block's strategy tests whether training on conversation data stripped of web-scraped noise produces an AI that talks more like a person and less like a search engine.
What do the sample conversations of Dessa reveal about its conversational abilities?
According to the article, the results of Dessa's conversations with humans land somewhere between impressive and unsettling, with the model occasionally engaging in back-and-forth exchanges that can feel awkward. This suggests the conversation-trained model produces outputs that feel more natural but sometimes uncanny compared to typical AI assistants.
Further Reading
- Papers with Code - Latest NLP Research - Papers with Code
- Hugging Face Daily Papers - Hugging Face
- ArXiv CS.CL (Computation and Language) - ArXiv
