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

Introduction

Block, the fintech company founded by Jack Dorsey, has unveiled Dessa, an experimental AI model that talks to you like an awkward new friend. Unlike most voice assistants that rely on text transcriptions, Dessa is a true speech‑to‑speech system trained from scratch on raw human conversation. The demo videos show a mix of charming banter, occasional missteps, and moments that feel surprisingly human. Block presents Dessa as a research prototype meant to explore how AI can learn language the way babies do—by listening, not by reading.

How Dessa Works

• Purely auditory training – The model’s knowledge comes exclusively from recorded speech. No text corpora, subtitles, or written scripts were used in its training data.
• End‑to‑end speech pipeline – Dessa converts incoming audio into an internal representation, generates a response, and synthesizes it back into speech, all without an intermediate text step.
• Local execution – According to the team, the model can run on a high‑end MacBook Pro (M4 Max or better), demonstrating that the architecture is lightweight enough for on‑device inference.
• Safety focus – By staying in the audio domain, the researchers argue that Dessa can better detect emotional cues such as pain or distress, addressing a common safety blind spot in text‑based assistants.

What the Demos Reveal

The public demos feature a variety of conversational snippets:

1. Guessing games – A voice‑only back‑and‑forth where the model misidentifies the speaker as “Ronaldo” before correcting itself, highlighting both its playful side and occasional confusion.
2. Film chatter – A discussion about Christopher Nolan’s movies that veers into an unexpected tangent about Bryan Singer’s Superman films, showing how the model can follow a thread but also wander.
3. Travel anecdotes – A user describes a surprise trip to Paris, and Dessa responds with a “teary gratitude” tone, illustrating its attempt to mirror emotional nuance.
4. AI‑to‑AI conversation – When paired with OpenAI’s GPT‑Live‑1 voice model, Dessa appears more fluid, suggesting that keeping the interaction in the audio domain may reduce latency and preserve natural timing.

These examples are deliberately raw; they expose the model’s quirks—awkward pauses, occasional misrecognitions, and moments of uncanny similarity to human speech. The team frames these imperfections as part of the research journey, akin to a toddler’s early attempts at conversation.

Why Block Is Building an Awkward Friend

The project is driven by three stated motivations:

• Curiosity (fun) – Exploring whether a system can acquire language purely from listening, without any written data.
• Human‑AI understanding – Improving the ability of AI to pick up on tone, pacing, and emotional cues that are lost when speech is flattened into text.
• Safety – Enabling voice agents to recognize distress or pain in a speaker’s voice, which could prevent harmful outcomes in critical applications.

Block plans to grant early access to a small group of AI‑safety researchers, hoping that external scrutiny will surface both promising directions and hidden risks.

Practical Takeaways

• For developers: Dessa demonstrates that end‑to‑end speech models are feasible on consumer hardware, encouraging experimentation with on‑device voice AI that respects privacy.
• For users: The current prototype is not a polished product; it’s a research showcase that may feel “awkward” but offers a glimpse of future voice assistants that truly listen.
• For the industry: If speech‑only training proves scalable, it could shift the AI landscape away from massive text datasets toward more privacy‑preserving, multimodal learning.

Conclusion

Block’s Dessa is an early‑stage experiment that trades polish for insight. By teaching an AI to converse solely from human speech, the team is probing the limits of how natural voice interaction can become—and whether that naturalness can improve safety and emotional awareness. While the conversations may still feel like chatting with an uneasy new friend, the underlying research could pave the way for the next generation of voice‑first AI.