Block has opened limited access to Dessa, a two-way speech model, for AI safety researchers, centering the project on vocal cues that text transcripts may fail to preserve.
Block says spoken exchanges can reveal signs of pain or distress that a text transcript may not capture. The company has opened access to the system to a limited set of researchers focused on AI safety, and its project page gives no public release date.
Block co-founder Jack Dorsey announced Dessa Oct. 23. He described the demonstrations as “full of emotion, nuance, pauses, and a natural cadence.” The company describes the model as full-duplex, meaning it can listen and deliver a spoken response at the same time.
Block says it trained the model from scratch on recorded human conversations, without written text or a pretrained model, using one NVIDIA GB200 NVL72 system. Its project page gives no public release timetable and does not disclose the model’s scale, training corpus volume or evaluation results.
Block has separately said Dessa runs faster than real time on a MacBook Pro M4 Max, a claim covered in EconoTimes’ earlier story about the model’s laptop performance.
A 2024 University of Illinois study examined a different system: voice-enabled GPT-4o agents with browser tools in simulated scams. The agents succeeded in 36% of tests overall and in 40% of five runs involving a crypto transfer. The results cover only the tested GPT-4o setup; they provide no finding on whether it could persuade victims or how Dessa performs.
