Guide

What to Test First When Your Dessa AI Invite Arrives: A Test Plan

2026-11-02 · 7 min read · By DessaAI Daily

The most-clicked page on this site is the waitlist guide — readers are not just curious about Dessa, they are queuing for it. So it is worth planning ahead: when an invite finally lands, your first session should not be ten random minutes of "say something funny." Dessa is a research preview, access will be limited at first, and the interesting questions about a speech-only, full-duplex model are specific. This is a ready-made test plan: what to try first, in what order, and what each test actually tells you — plus what there is no point testing, because Block has already been clear about the limits.

First, the honest framing. Block has not announced when invites will reach general waitlist members, or what form access will take. Safety researchers go first, by design. Our article on when Dessa AI will be available lays out that order, and the Dessa AI waitlist guide covers joining the queue and avoiding fake access offers. This plan assumes you have genuine access through Block — never through a third party selling "accounts."

Start with the thing Dessa claims to do differently: conversation itself

Most voice assistants turn your speech into text, think in text, then speak the reply — which throws away tone, timing and emotion. Dessa was trained from scratch on human conversation audio only, with no text in the middle. So test conversation, not trivia. Talk the way you would on a phone call: hesitate, change your mind mid-sentence, trail off. Does it wait naturally, or pounce on every pause? Does it backchannel — the "yeah," "got it" and "ooh" sounds Block's demos feature — at moments that feel right, or at random?

Then interrupt it, politely and then firmly, while it is mid-answer. Full-duplex means it listens while it talks; our plain-English explainer on what full-duplex means and the deep dive into how Dessa works explain why interruption handling is the real signature of this design. A model that stops cleanly, keeps its place and resumes without restarting from zero is doing something turn-based systems genuinely struggle with.

Test the emotion channel — in both directions

Block's central claim is that text throws away information Dessa keeps: emotion, sarcasm, laughter, sighs. So say the same sentence three ways — pleased, frustrated and joking — and see whether the replies differ in ways that fit. Try genuine sarcasm, the hardest case in the demos' own terms. Then test the outbound direction: tell it something surprising and something sad. In Block's recordings it gasps, laughs mid-sentence and can sound teary with gratitude. Your question is simpler and more useful: does the emotion sound responsive to what you said, or decorative?

Also try the untaught behaviour Block highlights: ask it to speak more quietly, then louder. Block says volume adjustment emerged without being taught. If it works in your session, note the exact wording you used — emergent behaviours are exactly what a research preview is for surfacing, and specific notes are worth more than impressions.

Probe the honest limits early, so they do not surprise you later

Do not waste your first session trying to catch Dessa out on facts and then concluding it is "bad." Block's own framing, and our article on what Dessa can't do yet, are clear: a model that learned language from speech alone, from hundreds of times fewer words than text models see, is not built to be a search engine, a reader or a tool-user. Ask it a factual question or two to calibrate — then move on. What you are testing is a conversation engine, and the fair questions are conversational ones: coherence over ten minutes, remembering what you said earlier in the same chat, finishing your sentence the way the demos show it can — and noticing when it gets awkward, as even Block's recordings occasionally do, including calling a speaker by the wrong name.

A simple first-session checklist

• Minutes 1–3: ordinary chat, with pauses, restarts and a mid-sentence change of mind.
• Minutes 4–6: interruptions — one gentle, one firm — while it is speaking at length.
• Minutes 7–9: the same line delivered happy, frustrated and sarcastic; one surprising piece of news; one sad one.
• Minutes 10–12: volume up, volume down; a laugh test — say something genuinely funny rather than asking it to laugh.
• Final minutes: one long thread — a story or a plan — to test memory and coherence across the whole conversation.

Write down what happened while it is fresh: which tests it passed, where it felt awkward, and the exact phrasing that triggered anything odd. If access comes with a feedback route to Block, that specificity is gold — a small lab iterating on a research preview can act on "it interrupted me twice during pauses at minute five" in a way it cannot act on "felt weird."

And if your invite has not arrived yet, you are not behind. The waitlist is the only public door, the demos on Block's official page are the same evidence everyone else has, and this plan will keep. When access widens — in whatever form Block announces — you will be ready to learn more in twelve minutes than a dozen random chats would teach you.

Sources: Block's official Dessa page (block.xyz/dessa), October 23, 2026 — demo behaviours, access order and limits as announced. Access timing for general waitlist members has not been announced.

Frequently asked questions

← More Dessa AI articles Join the official waitlist ↗