ഇന്റർനെറ്റില്ലാതെ പ്രവർത്തിക്കും; മനുഷ്യരെപ്പോലെ സംസാരിക്കുന്ന 'ഡെസ' എഐയുമായി ജാക്ക് ഡോർസിയുടെ ബ്ലോക്ക്

Jack Dorsey's Block has introduced Dessa, an AI model trained entirely on human speech without text or pretrained models. It can listen and talk at the same time, picking up emotions, sarcasm and natural pauses.

Jack Dorsey, the man behind Twitter (now X), is back with an AI model called Dessa that doesn't run on any LLM (large language model). Instead, it is trained through speech-to-speech human conversation. Most conventional voice AI models, including ChatGPT and Gemini, first turn your voice into text before replying. This could filter out a lot of information that text alone can't carry, such as emotions, sarcasm, laughter and the way people naturally pause while speaking.

Dessa takes a different route. Developed by a small AI research lab at Block, Dorsey's company, it learns language by listening to human speech rather than reading text. The team says it can hold natural, real-time conversations and runs locally on a MacBook Pro M4 Max or a better machine, without needing an internet connection. The team also released a range of samples, Dessa's conversation with humans on the website as well as on X, which took people by surprise.

Dessa can listen and talk at the same time

Unlike conventional voice assistants that wait for you to finish speaking before responding, Dessa can listen and talk at the same time. This is known as "full-duplex" speech-to-speech technology. It can interrupt naturally, respond with quick acknowledgements such as "yeah" and "got it", and even laugh in the middle of a sentence when something is funny.

The team says Dessa can recognise emotion, sarcasm and changes in tone that may get lost when speech is converted into text.

Dessa's creators revealed it was trained using a single NVIDIA GB200 NVL72 system. They also describe it as a new area for AI research, raising questions about how models can learn language and facts from sound alone.

Jack Dorsey says Dessa runs locally

Dorsey also answered a couple of questions about the model on X. When a user asked what language model powers Dessa's voice, he replied, "speech to speech. no llm." Another one asked whether it can run locally; Dorsey answered, "Yes".

The team is now inviting a small group of AI safety researchers to try Dessa and look for risks the developers may have missed. The model is an early research project, but its approach offers a different idea of how AI could learn to communicate: not by reading everything first, but by listening to how people speak.