Inside Thinking Machines’ Interaction Models

2026-06-30 · Source: ByteByteGo Newsletter · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Emerging Technologies & Innovation · Depth: Expert, extended

Summary

Thinking Machines, an AI research lab, proposes "interaction models" to overcome the limitations of current turn-based AI systems in real-time human-AI collaboration. They argue that existing voice AI, built on a "harness" of helper systems around a single-threaded language model, creates a bottleneck by forcing humans to adapt to the model's perception. Their solution, exemplified by TML-Interaction-Small, a 276-billion-parameter mixture-of-experts model with 12 billion active parameters, prioritizes continuous audio and video input/output. Key architectural innovations include time-aligned micro-turns, processing data in 200-millisecond chunks, and a two-model coordination scheme pairing a fast interaction model with a slower background model for deep reasoning. This approach enables new capabilities like speaking while listening or proactive visual interjections, validated by custom benchmarks such as TimeSpeak and RepCount-A, which existing models fail.

Key takeaway

For AI Architects designing next-generation interactive systems, recognize that current "harness" architectures impose fundamental limits on real-time human-AI collaboration. You should explore integrating continuous, time-aligned multimodal processing directly into your core models, potentially adopting a two-model coordination strategy for responsiveness and deep reasoning. This paradigm shift is essential to move beyond simulated conversations towards truly fluid, collaborative AI experiences.

Key insights

Integrating interactivity directly into AI models, rather than relying on external "harness" components, unlocks true real-time human-AI collaboration.

Principles

• Turn-based AI limits genuine human-AI collaboration.
• External helper systems create capability ceilings.
• Continuous multimodal streams enable fluid interaction.

Method

Design AI systems with time-aligned micro-turns (200ms) for continuous multimodal input/output, coordinating a fast interaction model for real-time response with a slower background model for deep reasoning.

In practice

• Enable live translation by speaking while listening.
• Support real-time visual feedback, e.g., counting pushups.
• Facilitate mid-sentence corrections based on context.

Topics

• Interaction Models
• Human-AI Collaboration
• Multimodal AI
• Real-time AI
• Mixture-of-Experts
• AI Architecture

Best for: Research Scientist, NLP Engineer, Computer Vision Engineer, AI Scientist, Machine Learning Engineer, AI Architect

Related on AIssential

• Interaction Models: A Scalable Approach to Human-AI Collaboration — Native, real-time multimodal interaction within AI models enhances human-AI collaboration beyond turn-based systems.
• Towards AI That Can Actually Interact — Real-time, continuous interaction models represent a fundamental shift from turn-based AI, enhancing human-AI collaboration.
• [NOTICE] This feed is no longer maintained — Native interaction models enable real-time, multimodal human-AI collaboration by integrating interactivity directly into the AI's core architecture.
• The Sequence AI of the Week #863: The Model is the Interface: Inside Thinking Machines' Interactive Models — Thinking Machines proposes "the model is the interface" for temporal, collaborative AI interactions, transcending sequential text processing.
• Thinking Machines drops a new, highly responsive model designed for humanlike interactions in real time — New AI architecture enables full-duplex, real-time multimodal interaction by processing data in 200-millisecond micro-turns.
• Mira Murati’s Thinking Machines Lab Introduces Interaction Models: A Native Multimodal Architecture for Real-Time Human-AI Collaboration — Native multimodal architectures can enable true real-time human-AI collaboration by integrating interactivity into model weights.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by ByteByteGo Newsletter.