Jack Dorsey’s Block made an AI model that talks to you like an awkward new friend

Jack Dorsey’s company Block has unveiled a new artificial intelligence model named Dessa, and the approach behind it is quite distinct from the text-heavy giants dominating the market. According to reporting by The Verge, Dessa was trained from scratch using nothing but human conversation data. This means the model does not rely on traditional text corpora but instead learns the nuances of spoken language directly from audio recordings of people talking.

The result is an AI that can engage in full-duplex speech interactions, allowing for interruptions and natural flow much like a real person. However, the early demonstrations reveal a model that is still finding its footing in social dynamics. Block shared chats that are mostly natural but occasionally stumble into awkward or uncanny territory, reminding us that speech is far more complex than simple question and answer exchanges.

One particularly telling example shared by Dessa lead Ragavan Thurairatnam shows the model initially mistaking the speaker for someone named Ronaldo. While this might seem like a simple error, it highlights the challenges of identity recognition and context retention in real-time voice processing. The back and forth remains largely coherent, yet these small glitches underscore the difficulty of replicating human social cues in a machine.

This development signals a significant shift in how tech leaders are approaching voice AI. Instead of just adding voice capabilities to existing text models, Block is attempting to build a foundation where speech is the primary mode of learning. This could lead to more intuitive assistants that understand tone, pause, and interruption better than current text-to-speech systems.

The implications for user experience are profound. If an AI can handle the messy, overlapping nature of human speech without breaking character, it could become a more viable companion for daily tasks. However, the current awkwardness suggests we are still in the early stages of mastering this technology. Users should expect a learning curve as these models refine their social awareness.

For developers and entrepreneurs, this is a clear signal that voice interfaces are moving beyond simple command execution. The race is no longer just about accuracy but about naturalness and social intelligence. Building tools that can navigate the subtleties of conversation will require new training data and architectural approaches that prioritize audio over text.

What this means for you is that voice-based AI is becoming more conversational but also more unpredictable. As you integrate these tools into your workflow, expect them to be better at handling interruptions but potentially weaker at maintaining long-term context. To test this yourself, try using a voice-enabled AI assistant to plan a complex multi-step project. Ask it to summarize the plan in real-time while you interrupt with corrections. This will help you gauge how well the model handles dynamic, human-like interaction versus rigid command structures.