Block Built a Talking Model That Never Saw Text

The page says every clip is a live recording and that nobody steers Dessa. Subtitles came later. In one, the model laughs while a person is still talking, which Block contrasts with systems that wait until you finish and then laugh on cue. In another, someone describes a surprise anniversary trip. Block says Dessa gasps "oh my god" under its breath and answers with what sounds like teary gratitude. Jay Peters at The Verge heard a mostly ordinary conversation that opens with the model deciding the speaker is a man named Ronaldo. There is also a clip of Dessa talking to OpenAI's GPT-Live-1. Block's disclaimer says both voices are generated, and that the developer of GPT-Live-1 has not reviewed or endorsed any of it.

Block says the model runs locally on a MacBook Pro with an M4 Max or better, and that training fit on a single NVIDIA GB200 NVL72 rack. The lab also says Dessa picked up language from hundreds of times fewer words than a large language model needs. That figure is theirs. Nobody outside the company has published a check on it. What Block has not published is a parameter count, a price, an API, or the weights. Ivan Mehta at TechCrunch, writing Saturday morning, said there is no public access yet. The page asks safety researchers to email ragavan+dessa@block.xyz. Ragavan Thurairatnam, who leads the lab, writes that a model trained on sound alone is new territory, and that he wants outside eyes on risks the team has not thought of. The group is small, it grew out of a Toronto deep-learning startup also called Dessa, and it reports to Dorsey.

I went looking for a warning about scam calls. The safety section on the page spends its time on a different scene. Block lists three reasons for the project: curiosity, a better understanding between people and AI, and safety. Then it asks you to picture someone in serious pain while a voice system has no idea and keeps going. Their view is that these systems are getting more power while staying deaf to how a person sounds, and that deafness will make them talk past you when you need them to stop. A voice that can cry on a laptop has an older problem sitting next to that one. Someone on the other end may believe the crying is a person. Block handles that problem by keeping the model off the public internet and inviting researchers in. It handles the deafness problem by arguing that Dessa can hear you.

I trust the engineering description more than that moral. A model with no text has to pick up facts from overheard talk, and the clips show it finishing a sentence and sounding moved by a trip it is not on. If the only public evidence is a set of demos the company chose, the empathy is a demo too. The bad calls are missing. The sigh in the wrong place. The model talking over someone who is hurt. The copy on a laptop telling a stranger it is sorry in a voice that cracks. Those are the recordings a safety researcher would ask to hear. A launch page is built to leave them off.

Block says, as far as it knows, Dessa is the first AI you can hold a realistic live conversation with after training only on human speech. "As far as we know" is carrying a lot of weight. So is the choice to keep the weights behind a researcher inbox while the clips travel on their own. You can already hear a voice crack. You cannot yet ask a question the company did not put on the site.