Home / Technology / AI's Leap: Beyond Turn-Based Chat
AI's Leap: Beyond Turn-Based Chat
12 May
Summary
- New AI interaction models treat interactivity as a core architectural element.
- Models process 200ms chunks of input/output simultaneously for full-duplex communication.
- Thinking Machines' AI achieves near-human response times on key benchmarks.
AI is shifting from "turn-based" interactions, where users input data and wait for a response, to more fluid real-time engagement. Thinking Machines has announced research preview models that treat interactivity as a fundamental part of their architecture rather than an add-on.
These new 'interaction models' feature a "full-duplex" design, processing 200ms chunks of input and output simultaneously. This allows AI to listen, speak, and perceive visual cues in real-time, enabling natural back-channeling and interjections.
On benchmarks like FD-bench, these models achieved a turn-taking latency of 0.40 seconds, significantly outperforming competitors such as Gemini and GPT. The models also showed strong proactive engagement in visual tasks.
This innovation holds significant potential for enterprises, enabling real-time monitoring in manufacturing, improving customer service response times, and managing time-sensitive processes more effectively.
Thinking Machines, founded by former OpenAI figures, has previously released a managed API for fine-tuning models called Tinker and secured substantial funding. The company has also been involved in significant compute partnerships.
