Thinking Machines previews real-time interaction models

This is a more interesting direction than yet another smarter chatbot: it treats interactivity as a first-class model capability, not a voice wrapper. If the latency and benchmark claims hold up outside the demo, this could reshape how AI assistants are built.

• –The full-duplex setup plus a separate background model is a clean architectural split: keep the conversation fluid, push heavier reasoning off-thread
• –Native interruption, backchanneling, and simultaneous tool use matter for real work far more than static benchmark gains
• –Training the interaction layer from scratch suggests this is infra-heavy research, not just prompt engineering or UI polish
• –The big risk is product usefulness: many users still prefer reliable turn-taking over a model that interrupts them well
• –If Thinking Machines can make this feel natural, it has a real shot at defining a new AI interface primitive