thinking machines beat openai to continuous ai interaction

thinking machines just shipped something the big labs haven't

while openai and google keep optimizing for longer autonomous runs, @thinkymachines is betting the future of ai looks more like a conversation than a job queue

their new interaction model – released as a research preview yesterday – ditches the turn-based paradigm entirely.

instead of waiting for you to finish, processing, then responding, it runs in continuous 200ms loops, taking in audio, video, and text all at once. it can interrupt you, react to what it sees on screen, speak while you're still talking, and track what's happening in the real world without being asked

every other realtime system today is a normal language model with voice-detection bolted on. thinking machines trained interactivity into the model from scratch – which means it gets better at collaboration the same way models get better at reasoning: scale.

for builders, the implications are immediate. proactive visual reactions, live translation, real-time error catching, ambient awareness – none of this requires custom scaffolding anymore. it's just what the model does

the preview is limited, and the 276B MoE model is too big to run locally. but the direction is clear.

ai interfaces are about to stop feeling like forms and start feeling like people