Dessa: The AI That Learned to Talk Without Reading
Dessa: The AI That Learned to Talk Without Reading
Block's new voice model learned conversation from sound alone — no text, no pretrained models. It laughs mid-sentence, cuts in when you pause, and runs on a laptop.
Jack Dorsey's Block unveiled Dessa on October 23, a voice AI trained from scratch on nothing but recorded human conversation — no text, no pretrained models — and Dorsey's announcement post has since passed 8 million views.
That's the whole pitch, and it's a strange one: an AI that never read a single word, yet can hold a conversation that sounds startlingly human. Block calls it "the most authentic full-duplex speech model ever built," and the demos are genuinely hard to ignore.
Your voice assistant reads. Dessa just listens.
Here's how the voice AI on your phone normally works. It turns your speech into text, runs that text through a language model, then turns the answer back into speech. Three steps, three places where timing and feeling get lost.
Dessa skips all of that. It's a speech-to-speech model, meaning audio goes in and audio comes out, with no transcription step in between. And it's full-duplex — a term borrowed from phone calls — meaning it listens and talks at the same time, the way people do.
That changes the rhythm completely. Dessa can cut in mid-sentence and slip in a "yeah" or "got it" while you're still talking. Block's argument is that text-based voice AI throws away everything that isn't a word — the pause, the sigh, the sarcasm — and that's why talking to it feels so stiff.
What the demos actually sound like
Block published around twenty recorded conversations on its Dessa page, and says nobody steered the model in any of them — Dessa picked what to say and how, down to its own voice. Subtitles were added afterwards.
The clips are the strongest part of the announcement. In one, Dessa laughs mid-sentence, the way a person does when something lands funny. In another, a speaker blanks on a movie title and Dessa finishes the sentence for them. In a guessing game posted by Dessa lead Ragavan Thurairatnam, it guesses the object on the first try — twice — then keeps playing after being told it already won.
There's also a demo where Dessa talks to GPT-Live-1, OpenAI's voice model, and the difference is the point: next to Dessa's natural cadence, the other model sounds robotic. Block notes GPT-Live-1's developer isn't affiliated with the project and didn't review it.
It isn't flawless, though. The Verge found the demos mostly natural but occasionally awkward — in one, Dessa mistakes the speaker for someone named "Ronaldo," and in another it drifts from Christopher Nolan's films into Bryan Singer's Superman movies mid-chat. Small slips, the kind a distracted friend makes on the phone. But they matter, because these are Block's own picked recordings, not an independent test.
Trained on one machine, runs on a laptop
The hardware story is almost as surprising as the voice. Block says it trained Dessa on a single NVIDIA GB200 NVL72 system — one rack-scale unit housing 72 Blackwell GPUs — and that the model learned language from hundreds of times fewer words than a text-based language model needs.
That claim needs a caveat: Block hasn't said which models it compared against, or how it counted. Still, the direction is striking. Speech is a much harder training signal than text — a spoken word takes thousands of times more bytes than the same word typed — yet Dessa apparently needed far less of it.
And once trained, it runs locally. Block says inference — the model generating its replies — runs faster than real time, and the whole thing runs on a MacBook Pro with an M4 Max chip or better, no cloud needed. If that holds up outside Block's lab, it's a very different shape for voice AI: private, offline, and on hardware you already own.
Why Block thinks this is about safety, not just vibes
Block's safety argument is blunt: today's voice AI is nearly deaf to emotion, and that's dangerous. Its example is a person in serious pain talking to a voice assistant that has no idea and just keeps going, because the text transcript stripped out the distress.
A model that hears tone, hesitation, and urgency natively could catch what a transcript misses. That's also why early access is going to AI safety researchers first — Block says it wants outside eyes on a model trained in a genuinely new way before anyone else gets it.
What you can't do with it yet
Here's the honest part: you can't try Dessa today. It's a research preview behind a waitlist, with early access limited to a small group of safety researchers. Block hasn't published a parameter count, an API, pricing, weights, or any benchmark scores.
So everything impressive about Dessa right now rests on Block's own demos. The voice sounds real and the local-on-a-laptop claim is testable — but until independent researchers get their hands on it, treat the superlatives as Block's, not settled fact.
Still, Dessa is worth watching for one reason: it proves a different path exists. Voice AI didn't have to go through text — it just always did, because text was easier. Block's bet is that listening first, the way babies learn to talk before they ever see a word, makes for a voice assistant that sounds like a person instead of a machine reading a script. If independent tests back that up, every voice assistant you use is about to change.
Sources: Block: Dessa, Dealroom, Notebookcheck, Gizmodo
