Block posted news of Dessa at 16:53Z on October 23, describing a full-duplex speech-to-speech model built from scratch on human conversation rather than fine-tuned from an existing text model. Three hours later, at 19:55Z, the same account opened an early-access call aimed specifically at AI safety researchers. No technical paper, benchmark, or demo accompanied either post.
The distinction Block is drawing matters to anyone who tracks how voice models get built. Most speech systems available today bolt audio capability onto a model that already learned language from text, then adapt it to handle turn-taking and interruption. Training a model from scratch on recorded human conversation is a different starting point, one that in principle could produce different timing behavior, since full-duplex systems have to manage overlapping speech and interruptions in ways text-trained models often handle poorly.
What Block has actually shown so far is an announcement and a recruitment call, not a working demonstration or a dataset description. The post does not disclose how large the training corpus was, what counts as human conversation in its sourcing, or how full-duplex behavior was measured, if it was measured at all. Block did not disclose specifics on model size, training compute, or evaluation methodology. That absence is not unusual for a same-day product announcement, but it means every claim about performance currently rests on the company's own framing.
For researchers deciding whether to request access, the practical checks are narrow but concrete: what conversational data was used and how it was collected, what counts as a full-duplex evaluation in this context, and whether safety researchers get raw model access or a constrained interface. The timing of the access call, just over three hours after the public announcement, suggests Block wants outside scrutiny before wider release, which is a reasonable posture for a conversational model trained on human speech data at scale. Whether that scrutiny shapes the eventual public version is unknown.
The open question is whether training from scratch on conversation actually produces different behavior than adapting a text model, or whether the distinction mostly affects marketing rather than measurable interaction quality. Until Block publishes details on data, evaluation, or safety findings from the researcher program, independent verification is not possible. The early-access window is the only mechanism currently available for anyone outside the company to test that gap.