Post #6895

🧠 Thinking Machines is replacing turn-based AI with always-present AI. They just unveiled TML-Interaction-Small, a 276B-parameter MoE model built for continuous real-time interaction instead of stop-start conversations. Most AI voice systems still work like walkie-talkies: you talk → they stop listening → they answer. Thinking Machines slices audio, video, and text into 200ms micro-turns, letting the model: • listen while speaking • react to gestures and screen changes • interrupt naturally when needed • track silence and timing • call tools in the background without leaving the conversation The key difference is architectural: interactivity is trained into the model itself instead of patched together with separate voice detectors, speech systems, and timing rules. Early results are impressive: • 0.40s turn-taking latency • 77.8 FD-Bench interaction score • 43.4% on Audio MultiChallenge This feels less like a chatbot waiting for prompts and more like AI that stays continuously present during interaction. Source. @aipost🏴