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

Block has introduced Dessa, an AI model it says was trained from scratch on nothing but human conversation. It's a true speech-to-speech system: sound in, sound out, with no detour through written text. Block says it can run locally on a MacBook Pro M4 Max or better.

Most voice assistants today turn speech into text, think in text, and read an answer back. That loses timing, hesitation and emotion. Dessa is a bet that a model that learns from speech alone can keep those things, and Block is framing that partly as a safety issue.

We find this genuinely interesting, and the reason is the method more than the demos. The team's pitch is that babies learn to talk without ever seeing a written word, so why not a model? Their answer is that it works, but it's very hard. That candor is welcome. Plenty of voice products sound polished because they're reading from a script. Dessa sounds like it's figuring things out as it goes.

The sample conversations show both sides. There's real back-and-forth with natural rhythm. There's also the strange stuff: the model mistaking a speaker for someone named Ronaldo, wandering off on a tangent about Bryan Singer's Superman films during a chat about Christopher Nolan, and responding to news of a surprise Paris trip with what sounds like teary gratitude. In a side-by-side with OpenAI's GPT-Live-1 voice model, Dessa comes across as the more natural speaker. Awkward, yes. But awkward like a person, not like a phone tree.

The safety argument deserves a hearing. Block's team argues that AI systems keep gaining power and independence while staying nearly deaf to emotion in our voices, and imagines a person in serious pain talking to a voice AI that just keeps going. A model that can actually hear distress could be a real improvement for support lines, accessibility tools and anywhere people talk to machines when things are going badly. We like that Block is giving a small group of AI safety researchers early access rather than rushing it out.

Now the downside. A model built to sound emotionally present, and to read emotion back, is also a model that can be very good at persuasion and attachment. Teary gratitude is touching in a demo. At scale, in products built to maximize engagement or sell things, it's a lever. Running locally helps on privacy, since conversations don't have to leave the device. It doesn't answer the harder question of what a voice that sounds like a friend should be allowed to do with that trust.

Our read: a thoughtful research direction with real promise for humane voice interfaces. I'm watching for what Block does with it commercially, because the line between empathetic and manipulative is thin when you're talking to the listener's emotions.

Block's team lists three reasons for building Dessa: curiosity, better understanding between humans and AI, and safety. Its argument is that speech is the most natural interface to any intelligence and that today's voice AI wastes it by converting what we say into text and failing to keep up with human timing. Dessa's effort is led by Ragavan Thurairatnam.