Review

Dessa AI Review: What We Know So Far (And What Nobody Can Test Yet)

2026-11-07 · 7 min read · By DessaAI Daily

Every review of Dessa AI has to start with an admission most reviews bury: nobody outside Block and its invited safety researchers can actually use it yet. There is no product to benchmark, no app to install and no price to judge value against. So this is a review of what genuinely exists in November 2026 — Block's official demos and claims, the reporting around them, and the long list of things that remain unproven — written so you can tell the difference between evidence and excitement. For the plain explainer version, start with what Dessa AI is; this article is the critical companion to it.

What we are reviewing, exactly

Dessa is Block's full-duplex, speech-to-speech AI model, announced October 23, 2026. Block trained it from scratch on human conversation audio only — no text, no pretrained models — and calls it, as far as it knows, the world's first realistic live-conversation model learned purely from speech. It hears and produces emotion, sarcasm, laughter and sighs; it backchannels while you talk; it knows when to interrupt; and its inference runs faster than real time locally on a MacBook Pro with an M4 Max chip or better. The review evidence is Block's published live recordings — unsteered, Block says, with the model choosing what to say, when and how — plus coverage by The Verge, TechCrunch and others. There are no independent hands-on evaluations yet. That caveat shapes every score below.

Conversation feel: the demos are genuinely different

On the one dimension the demos can demonstrate — how a conversation feels — Dessa is impressive in a way spec sheets cannot capture. It laughs the moment something is funny, mid-sentence, instead of waiting for a polite gap. It gasps "oh my god" under its breath before answering with what sounds like tearful gratitude. It finishes a speaker's sentence when they blank on a movie title. In our breakdown of what the official recordings show, the pattern is consistent: the timing is the technology. Because Dessa listens while speaking — the full-duplex design explained in how Dessa AI works — reactions land inside the conversation rather than after it. Verdict on feel, based only on curated recordings: the strongest public demonstration of natural voice timing so far, with the asterisk that Block chose the tapes.

Intelligence and knowledge: unproven by design

Here the review turns honest. A model that has never read a book, a manual or a webpage — by design — starts with a knowledge handicap against text-trained rivals, and Block has published nothing to measure it: no benchmarks, no parameter count, no training-data hours, no factual-accuracy testing. The demos show fluency, social awareness and a surprising amount of knowledge Block itself notes came from raw sound alone. They do not show Dessa writing code, summarising documents, using tools or answering checkable questions under pressure. Reviewers also caught very human awkwardness — in one recording it called a speaker "Ronaldo" — which is charming in a demo and would be costly in a work assistant. On knowledge work, the only honest score today is "not yet graded", and anyone who tells you otherwise is guessing.

Efficiency claims: remarkable, and still claims

Two numbers anchor Block's engineering story: Dessa learned language from hundreds of times fewer words than text large language models need, and it was trained on a single NVIDIA GB200 NVL72 system. If independent researchers confirm efficiency anywhere near that, it matters far beyond voice — it would suggest speech carries far more learnable signal per word than the field assumed. But today both numbers are Block's own, unreplicated, and published without the training details that would let outsiders check them. File them as "extraordinary, pending verification" — the same status as lead researcher Ragavan Thurairatnam's description of the launch as "a GPT-3 moment for speech".

Safety and the rollout: the most credible part

Paradoxically, the aspect of Dessa that inspires the most confidence is not a capability — it is a decision. Block is giving early access first to a small group of AI safety researchers, using institutional affiliation as a legitimacy signal, before any wider release. For a payments company whose earlier Toronto team became famous for a 2019 synthetic-voice demonstration, that caution reads as earned rather than theatrical; we covered the reasoning in is Dessa AI safe? A human-sounding, interruptible voice is a genuinely new impersonation surface, and Block is behaving as if it knows that. The flip side: no safety findings, safeguards or disclosure rules have been published yet, so the process is promising and its results are still to come.

Availability and value: there is no product to buy

A review normally ends with price and a buy-or-wait verdict. Dessa has no price, no API, no download, no weights release and no public release date — all officially "not announced", and our Dessa AI Tracker keeps that status board current daily. Joining the waitlist at block.xyz/dessa is free; anyone charging for access is not Block. So the practical verdict is simple. As a research result, Dessa is the most interesting voice AI story of the year: a credible team, a radical method and demos that make today's assistants sound like walkie-talkies. As a product, it does not exist yet — you cannot use it, build on it or budget for it. The review that matters will be written by the safety researchers and first independent testers. When they publish, this page's provisional judgements get replaced by measured ones.

Sources: Block's official Dessa page (block.xyz/dessa), including demo notes and team details; The Verge and TechCrunch coverage, October 2026. No independent evaluation existed at the time of writing.

Frequently asked questions

← More Dessa AI articles Join the official waitlist ↗