In one of Block’s Dessa demonstrations, a man announces a surprise trip to Paris. The AI gasps, then responds with what sounds like tearful gratitude.

That is the performance. Here is the detail I find more interesting: Block says this conversation model runs locally on a MacBook Pro with an M4 Max chip or better. The announcement and recordings

Imagine that becoming something you can run on your own laptop: a voice you can interrupt, practise a conversation with, or build into an app.

There is a catch. You cannot generally download Dessa yet. And sounding more human does not tell us how useful it will be.

I am Alex. Welcome to ShortCu8 by Innov8.

Lets Dive Deep.

Dessa comes from a research lab at Block, Jack Dorsey’s company. Block says it trained the model from scratch on human conversation, without text or pretrained models. It listens while speaking, and the selected recordings include laughter, interruptions and changes in volume. Block’s explanation

Think about saying “fine” after a good meal, then saying it after an argument. A transcript gives you the same word. Your voice carries the difference.

That is why speech is an interesting thing to learn from directly. A good conversation depends on knowing when to stay quiet as much as knowing what to say.

I would want to try a messy conversation myself before calling it a better assistant.

Yes, much of the conversational behaviour is already part of GPT Live.

OpenAI’s voice model can listen and speak simultaneously and handle interruptions. It can also send harder questions to a separate model for reasoning or search while keeping the conversation going. Developers can connect it to a backend that uses tools. How GPT Live works

GPT Live gives us a useful reference point: natural conversation connected to deeper work. We should ask Dessa to prove both parts.

Block even includes a recording of Dessa talking with GPT-Live-1. That is a demonstration, not a controlled test showing which model is better. The two-model conversation

For something you can use today, ChatGPT Voice is available subject to account and plan limits. A natural-sounding clip cannot settle questions about accuracy, languages or everyday reliability. ChatGPT Voice availability

“Runs on a Mac” needs a little unpacking.

Block specifies an M4 Max MacBook Pro or better. Its announcement does not give an exact parameter count, download size or minimum memory requirement. We can say it is claimed to fit on that machine. We cannot say it works on any MacBook. Hardware claims and missing details

Running it also differs from training it. Block says training used an NVIDIA GB200 NVL72 system. The laptop claim concerns using the finished model. Training details

If local operation holds up, developers could have more control over where conversations are processed. Keeping audio on-device could be useful for privacy, but the whole app would need to respect that choice. Add an online search service or cloud logging and some information may still leave the machine.

I like the possibility of having that choice. I would still check the actual app before calling a conversation private.

The first thing I would try is interview practice. Give the AI a role, answer aloud, get interrupted and try again. That could be more useful than rehearsing against a silent list of questions.

Conversation practice is another possibility, provided the model supports the language well. We do not yet have enough public evidence to promise good Malayalam or any particular accent.

For a business, imagine a voice front end for booking appointments or checking an order. The voice model handles the exchange; separate software needs to look up availability, confirm details and save the booking. GPT Live’s documented backend approach already shows how that division can work. Developer guide

These are possible applications for Dessa, not products Block has shipped. A model that chats convincingly does not automatically know your calendar or have permission to change it.

The useful distinction is between three questions: does it sound natural, does it know the answer, and can it carry out the job?

I would test all three before trusting a voice assistant with important work. A friendly response is pleasant. A correctly completed booking is useful. We should not confuse them.

For now, Block is inviting a small group of safety researchers to explore Dessa. Research preview

My interest is in what happens after those demos. Let people run it, interrupt it, ask awkward questions and build something useful around it.

That will tell us more than another clip of an AI sounding impressed.

We read every reply - just reply to this email and let us know how we can improve !

#AD1

80% of product professionals say AI helps them ship faster, but customers aren’t seeing value any sooner. Atlassian’s State of Product 2027 explores the gains AI is delivering and the challenges that remain, from decision-making that hasn’t kept pace to gut instinct overriding customer evidence.

Read the report

#AD2

If you have paid time left on an individual Granola plan, we'll match it with a Wispr Flow subscription that includes Notetaker and dictation, and add bonus time, up to 12 months total. Sign in or create a Wispr account and submit proof of your plan to check eligibility.

Try Notetaker Free