Thinking Machines Lab: Interaction Models for real-time human-AI chat
Mira Murati's lab dropped TML-Interaction-Small, a 276B parameter model that processes audio, video, and text in parallel 200ms chunks without external voice detection.
Key shift: two-track architecture runs a real-time interaction stream alongside a background reasoning engine, both seeing the full conversation context — eliminates the freeze-during-generation problem of turn-based systems.