The next natural level of Ai is speech. I personally know some companies that ar
By Brett Bohannon · October 24, 2026 · Curated by George's Blog
The next natural level of Ai is speech.
I personally know some companies that are implementing it and I look forward to this.
Dessa, was just announced by Block, below is a summary of it.
Block's small AI lab (the former Dessa team, reports to Jack Dorsey) released Dessa, a full-duplex speech-to-speech model trained only on human speech. No text, no pretrained LLM.
A few things stood out:
1. It listens and talks at the same time
2. It backchannels (yeah, got it) and laughs mid-sentence
3. It hears emotion, sarcasm, and sighs
4. It learned language from hundreds of times fewer words than LLMs
5. Trained on a single NVIDIA GB200 NVL72, and it runs faster than real time, locally on a MacBook Pro M4 Max
Most voice AI converts speech to text and throws the rest away. Block also framed this as a safety issue: AI that can't hear someone is in distress. Early access is opening to AI safety researchers.
My angle: text is a lossy layer. Same lesson applies to commerce agents. The closer the agent sits to the real signal, the better it works.