Models#AI safety#Model training#Model inference
Block unveils Dessa: a speech-to-speech model it says was trained from scratch on conversation alone, running locally on an M4 Max
Block's Dessa, released October 23: a full-duplex speech model trained on raw conversation with no text pipeline, per its makers, running locally on an M4 Max.
Dessa is a full-duplex speech-to-speech model announced on October 23 by Block, the payments company co-founded by Jack Dorsey. Its makers say it was trained from scratch on nothing but human conversation — no text, no pretrained models — that it runs faster than real time on a MacBook Pro with an M4 Max or better, and that all training fit on a single NVIDIA GB200 NVL72. There is nothing to download and no API: the site takes waitlist signups, and early access is reserved for a small group of AI safety researchers.
Key points
- Training route: trained from scratch on human conversation alone, per the official page, with a single NVIDIA GB200 NVL72 as the training machine.
- Interaction: full duplex — it listens and talks at the same time, interrupts, and backchannels; Block says it handles emotion, sarcasm, laughter and sighs natively.
- Availability: no weights, no paper, no API, no pricing; a public waitlist, with early access going to a small group of AI safety researchers.
- Reception: polarized on Hacker News (submitted by dsr12 as “Dessa: Conversation Model trained only on speech”) — praise for conversational realism, open-weights demands from critics.
Background
The standard recipe for voice assistants is a relay: speech-to-text converts your voice, a language model works on the transcript, text-to-speech reads the reply back. Both ends of that relay have shown up here before — Microsoft’s streaming transcriber MAI-Transcribe-2 on the way in, ElevenLabs’ v4 Turbo with its roughly 100 ms synthesis latency on the way out. The relay’s cost is what the transcript drops: sarcasm, hesitation, a cracking voice are gone once speech becomes text, and the system waits for you to finish before it starts. Full duplex has reached video first, in Tavus’ Griffin, which watches and talks at once; on the voice side, OpenAI’s real-time GPT-Live-1 is the best-known incumbent.
The name has history. Block’s lab is named after Dessa, a Toronto deep-learning startup, and the team is that startup’s old roster; lead Ragavan Thurairatnam founded the original and sold it. The group reports directly to Jack Dorsey. For Block it is another bet on AI — goose, its internal coding agent, is the company’s other public AI project.
The facts
Everything in this table is what Block’s page (block.xyz/dessa) states. The page and The Verge’s report went up the same day, October 23, 2026; we read both on October 24.
| Item | Official claim |
|---|---|
| Training data | Human conversation recordings only, trained from scratch: “no text, no pretrained models” |
| Data efficiency | Language learned from “hundreds of times fewer words” than LLMs need, with a joke that LLMs also learn code |
| Training hardware | One NVIDIA GB200 NVL72 |
| Inference speed | Faster than real time, no wait between turns |
| Local runs | A MacBook Pro with M4 Max or better |
| Interaction | Full-duplex speech-to-speech: simultaneous listening and talking, interruptions, backchannels, native handling of emotion and laughter |
| Availability | Waitlist on the site; early access for “a small group of AI safety researchers”; no API, no pricing |
Block calls Dessa “the world’s first AI you can have a realistic live conversation with that was trained only from human speech.” That “world’s first” is the company’s own; we found no independent confirmation. The page carries a main list of six full recordings — talking about the release, “Remembering a movie” (a user forgets the title, Dessa finishes it), “An anniversary surprise,” an expressive conversation, life advice, and the conversation with GPT-Live-1 — plus additional samples including one where Dessa mistakes the speaker for Ronaldo; the page notes every clip is an unsteered live recording with subtitles added afterward. Per The Verge’s description, the site’s examples also include a chat about Christopher Nolan that drifts into the Superman films and a teary-sounding expression of gratitude over a surprise trip to Paris; the guessing game is a video lead Ragavan Thurairatnam posted on X. The GPT-Live-1 side-by-side sits on the page, and the verdict that Dessa sounds more natural comes from Dessa’s own side; the site notes that GPT-Live-1’s developer is unaffiliated and has not endorsed anything.
The research team lists six full members — Louay Hazami, Hashiam Kadhim, Qi (Alex) Liu, Rayhane Mama, Aydin Polat, Marc Tyndel — plus interns. Thurairatnam’s launch post on X (posted October 23 at 4:38 PM) called Dessa “a GPT3 moment for speech”; when we checked on October 24, it had passed one million views.
What others say
The Verge’s reporter listened to the demos and called them awkward and uncanny; in one recording, Dessa mistakes the speaker for someone named Ronaldo.
On Hacker News, the most-upvoted and the harshest comments are about the same recordings. The positive side: “like talking to a friend on the phone,” speech realism that is “very good.” The criticism is more specific. IshKebab found the voice convincing but “the intelligence is pretty clearly 2024-level.” paxy objected that the model seems trained “to interrupt as much as possible,” leaving the human unable to finish a thought. ajam1507 called it a parlor trick mimicking “informal telephone mannerisms.” stanfordkid asked where intelligence is supposed to come from when the diet is speech corpora — is it “just generating plausible slop”? freeone3000 doubted the efficiency carries over to general-purpose text models. ilaksh asked whether it is built on Mochi or PersonaPlex; nobody could say. lrvick’s demand was the shortest: “Open weights or it didn’t happen.”
Abuse is the other thread. Several commenters expect scam calls aimed at the elderly to get harder to spot once fraudsters have voices that sigh and hesitate; others see AI companionship as the obvious market, with one invoking the film Her. None of this can be checked independently right now — the model is closed, and everyone, The Verge’s reporter included, has heard only the officially chosen recordings.
Our take
Against other entries in the speech-to-speech file, Dessa’s unusual part is the training claim. “More natural” is what every voice model says about itself; the checkable specifics here are a full training run on one GB200 NVL72 and language learned from hundreds of times fewer words than an LLM needs. If both hold up, the training budget for conversational speech models needs recalculating, and the old question of whether intelligence can grow from speech alone gains a data point.
Right now there is nothing to test. No paper, no weights, no third-party evaluation; HN commenters suspect the demos were cherry-picked, and The Verge’s listening notes show the gap between the recordings and the pitch. If you build voice agents, the pragmatic read is to treat de-texted speech models as a route big companies are now officially betting on, while shipping on the relay pipeline you can actually call from an API. Wait for the safety researchers’ first public reports — or an actual weights release — before rewriting any roadmap.
How to try it
Two moves are available: listen to the recordings and join the waitlist at block.xyz/dessa, and, if you are an AI safety researcher, apply through the form on the page — it asks for a name, an institutional email and, optionally, a Google Scholar or blog link. There is no model to download and no API to call.
What to watch
Four things: whether a technical report or paper appears; whether weights are released, the loudest ask on HN; what the safety researchers in early access publish first; and how incumbents respond — whether OpenAI, behind GPT-Live-1, engages with the no-text route at all.