 |
| Block's Dessa AI model learns language entirely through spoken conversations |
In a move that could fundamentally reshape how we think about artificial intelligence and language, Jack Dorsey's Block has announced Dessa—a full-duplex speech model that was trained entirely on human conversations without a single line of text. The announcement, shared by Dorsey himself on X (formerly Twitter), has sparked intense discussion among AI researchers and tech enthusiasts alike.
A Different Kind of AI
According to Block's official announcement, Dessa represents a fundamental departure from conventional AI development. While most language models are trained on massive text datasets scraped from the internet, Dessa learned everything it knows by listening to actual human speech—complete with pauses, laughter, interruptions, and all the messy nuances that text simply cannot capture.
"Audio in, audio out," the company explains. "Dessa is a true speech-to-speech model, and everything it knows came from human speech."
The model is described as "full-duplex," meaning it can listen and speak simultaneously, just like humans do. It can interrupt mid-conversation, acknowledge what someone is saying with natural backchannels like "yeah" or "got it," and even finish sentences when someone struggles to recall a word.
The Technology Behind the Breakthrough
What makes Dessa particularly remarkable is its efficiency. Block claims the model was trained using just one NVIDIA GB200 NVL72 system, which houses 72 Blackwell GPUs. Even more impressive, the resulting model can reportedly run locally on a MacBook Pro equipped with Apple's M4 Max chip or better—with inference running faster than real time.
Block's research team, composed of former members of the Toronto-based deep learning startup Dessa, has been building generative speech models for over eight years. As detailed on their research page, they believe voice is "the most natural way to interact with intelligence."
Why Speech Matters
The company argues that current voice AI systems are fundamentally flawed because they convert speech to text and back again, losing crucial information in the process.
"The most powerful AIs in the world can't hear how you feel," Block notes. "We think that's ridiculous, leads to a worse experience, and is potentially dangerous."
Dessa natively understands emotion, sarcasm, laughter, and even a "frustrated sigh or a voice cracking with sadness"—things that would be stripped away in traditional speech-to-text-to-speech pipelines.
Striking Demonstrations
Block has published several demonstration recordings showing Dessa in action. In one clip, the AI laughs mid-sentence when something is funny, rather than waiting for the speaker to finish and then laughing on cue. In another, it gasps "oh my god" under its breath during a story about an anniversary surprise, then responds with what sounds like genuine emotional engagement.
Perhaps most notably, Block shared a recording of Dessa conversing with OpenAI's GPT-Live-1 voice model—a demonstration that shows two AIs engaging in natural-sounding dialogue.
Safety First, Access Limited
Despite the excitement, Dessa isn't available to the public yet. Block is taking a measured approach, inviting a small group of AI safety researchers to evaluate the model first.
"A model that learned to talk from sound alone is new territory, and we want outside eyes on it," explained Ragavan Thurairatnam, Dessa's lead researcher. "We're giving a small group of AI safety researchers early access to find the risks we haven't anticipated."
The company has also not disclosed Dessa's parameter count, model weights, or standardized benchmark results—details that the research community will undoubtedly want as the model becomes more widely available.
The Road Ahead
In his announcement on X, Dorsey emphasized the natural quality of interactions with Dessa: "Talking with it feels like talking with a person…full of emotion, nuance, pauses, and a natural cadence."
Block has promised more recordings will be shared on X soon, and launch links for the model are reportedly coming in the near future. For now, those interested in following Dessa's development can join a waitlist through Block's website.
The question remains whether Dessa's speech-only approach will prove to be a genuine breakthrough or an interesting research curiosity. What's clear is that Block has opened a new frontier in AI development—one where machines learn to communicate not by reading, but by truly listening.
Sources: Block/Dessa, Jack Dorsey on X