Jack Dorsey's company builds an AI that talks without reading

Block has released an AI model called Dessa that it claims learned language without ever processing a single text file. The company, led by Jack Dorsey, describes the system as a true speech-to-speech model. It was trained from scratch on nothing but human conversation. This approach aims to fix a common flaw in current voice assistants, which often flatten natural speech into text and struggle to match human timing.

Ragavan Thurairatnam, who leads the project, shared one sample exchange to demonstrate the model's capabilities. In that recording, the conversation sounded mostly natural, although Dessa initially mistook the speaker for someone named Ronaldo. The company's website hosts additional examples. One clip features a discussion about Christopher Nolan films that drifts into a tangent about Bryan Singer’s Superman movies. Another shows a user telling the model about a surprise trip to Paris. The site notes that Dessa responded with what sounds like teary gratitude. A third clip pairs Dessa with OpenAI's GPT-Live-1 voice model, resulting in a more natural sounding exchange.

The development team lists three main reasons for building Dessa: curiosity, better human-AI understanding, and safety. They compare the learning process to human infants, noting that babies learn to talk without seeing written words. The team wanted to see if a model could do the same. They admit the process is very hard but successful. They also argue that current voice AI fails to capture the emotion in human speech. The team describes a scenario where a person is in serious pain, but a standard voice AI remains unaware and continues its task. They position Dessa as a step toward addressing that gap.

Dessa is designed to run locally on a MacBook Pro with an M4 Max chip or better. This allows the model to operate without sending audio to a cloud server. Block has not announced a wider release schedule for the general public. Instead, the company plans to give early access to a small group of AI safety researchers. These researchers will evaluate the model's performance and its ability to interpret emotional cues in voice.

The announcement highlights a shift in how AI companies approach voice interfaces. By skipping the text intermediate step, Block aims to preserve the timing and emotional weight of spoken language. The published recordings show a system that is still learning, with some interactions appearing slightly awkward or uncanny. However, the team maintains that the model represents a new path for voice-based intelligence. The focus remains on making the interaction feel more like a conversation between two people rather than a query-and-response cycle.