Block's New AI Model Trained on 0 Words of Text

Block has released a new artificial intelligence model named Dessa.

The company led by Jack Dorsey states the system was built using only human conversation. It did not use any written text during its training process. This approach mirrors how infants acquire language skills.

Ragavan Thurairatnam leads the project and shared sample audio clips. One recording shows the model interacting naturally with a user. However, the system initially mistakes the speaker for a person named Ronaldo.

The model can run locally on a MacBook Pro M4 Max or better hardware. The team says this setup allows the system to function without constant cloud connectivity. They describe it as a true speech-to-speech architecture.

The developers cite three reasons for creating the model. They list curiosity, improved human-AI understanding, and safety as the core drivers. They compare the learning process to a baby hearing speech without reading.

The team argues that current voice assistants lose data by converting speech to text. They claim this process flattens natural timing and nuance. Dessa is designed to listen directly to the audio stream.

The safety focus centers on emotional detection. The team notes that many AI systems ignore the emotional tone in a human voice. They propose Dessa as a step toward recognizing pain or distress in real-time conversations.

Block will provide early access to a small group of AI safety researchers. The company did not announce a timeline for a wider public release.