Jack Dorsey launches native speech model Dessa, which can run full-duplex conversations locally on an M4 Max MacBook

Holy shit, holy shit, this really deserves a holy shit, brothers. Twitter co-founder Jack Dorsey has pulled out a native speech model developed in secret, which can run locally directly on an M4 Max MacBook and easily handle full-duplex conversation just like a real person.

Damn, I feel like the current AI architecture has been completely destroyed. It doesn't rely on text, has no text large model, and is trained end-to-end purely on human audio, learning to listen and speak like a baby.

And there is no text intermediary, no need to call a cloud API, and it runs locally on an M4 Max, handling full-duplex conversation just like a real person.

I think this may be announcing that the crude era of first converting to text and then synthesizing speech is officially over.

Many people think this is just another more realistic-sounding, better-sounding AI voice toy made in Silicon Valley. But if you dig into its underlying architecture and take a careful look, you'll find it has done something extremely rebellious:

It directly kicked away the text foundation that today's large models treat as gospel!!!

Today's voice assistants, whether GPT-4o or the various real-time voice systems on the market, are essentially all three-stage assembled vehicles:

They convert your words into text (ASR), throw it to a text large model for reasoning, and then synthesize the text into speech and spit it back to you (TTS).

This approach castrates everything most precious in human language. The sneer, hesitation, sigh, choking up, and subtle irony in your words are already dead the moment they are translated into cold text characters.

Dessa took a completely opposite hardcore unorthodox path: Zero text, zero text-pretrained large model, trained end-to-end purely on real human conversation audio.

It doesn't know at all that something called text exists in the world. Like a human baby, it builds cognition directly from the sound signals of listening and speaking.

Precisely because it completely bypasses the text intermediary, it demonstrates native full-duplex capability that none of the current voice AIs possess. It can listen and speak at the same time, you can interrupt at any moment, it will laugh out loud halfway through a sentence when something is funny, and it can precisely catch the sigh in your tone.

The official data even slaps the current brute-force compute school in the face: The amount of tokens consumed in its training is several hundred times less than that of large language models. The entire training used only one NVIDIA GB200 NVL72, and it can run locally offline at faster-than-real-time speed on a Mac laptop equipped with the M4 Max chip.

We used to always think that large language models were the master key to all intelligence.

But this experiment has ruthlessly torn open a blind spot: Text is merely a low-bandwidth compression tool invented by humans to overcome time constraints, while sound and physical signals are the high-bandwidth essence of intelligent interaction.

Of course, amid the excitement, let's keep our cool as usual, haha. This is currently only a small-scale research preview from Block's internal lab, and the demos on the market have all been carefully selected by the official team.

Away from the vast text knowledge base, how long a model trained purely on audio can hold up in complex multi-step logical reasoning is still a huge question mark. But I think at least it has released a highly penetrating signal: While everyone is using tens of thousands of GPUs to extend text context indefinitely, the most fundamental revolution in intelligent interaction is already launching a major counterattack toward humanity's most primitive auditory instinct!