Thinking Machines released 'Interaction Models', their first AI model after two years and $2B in capital. Rather than competing with frontier models from OpenAI or Anthropic, they focus on real-time conversational interaction using a fully-duplex voice system. Key features include micro-turn switching (200ms chunks of listen/speak), delegation of hard reasoning tasks to a background smart model, video input for reading facial expressions, and a significantly larger model size than existing fully-duplex systems like Moshi. The author views the scale and video capabilities as genuinely novel, while noting the fully-duplex interaction features themselves are not new. The reasoning model integration is seen as both a legitimate architectural choice and a potential benchmark-boosting tactic.
Sort: