Microsoft releases MAI Realtime full‑duplex speech model as GPT‑4 shown to predict human behavior

Microsoft’s MAI Playground now includes a hidden early‑access entry called MAI Realtime, a full‑duplex speech model that can listen and speak simultaneously. The system supports dozens of languages, offers two synthetic voices (Victoria and Grant), and provides a debug panel showing latency and processing steps. No official announcement has been made, and the model operates without a dedicated public release.

A separate study published in Nature examined whether large language models can forecast human actions. Researchers fed GPT‑4 the design of 70 U.S. survey experiments involving more than 119,000 participants and asked it to simulate responses. The model’s predictions closely matched the actual outcomes, though it tended to double‑estimate effect sizes, a bias also seen in human forecasts. Accuracy dropped when the model attempted to predict results from 15 field experiments.

Both pieces illustrate rapid advances in AI: Microsoft’s real‑time conversational capabilities and GPT‑4’s emerging role in social‑science prediction.