SourceStalecollected in 12m

Mira Murati's Firm Teases Interaction Models

Read original on The Verge
#multimodal#real-time-ai#human-ai-interaction

Ex-OpenAI CTO's new AI co. unveils real-time multimodal interaction models for natural collab

30-Second TL;DR

What Changed

Mira Murati launches Thinking Machines post-OpenAI

Why It Matters

This advances multimodal AI towards more human-like interactions, potentially disrupting real-time applications like virtual assistants. Ex-OpenAI leadership signals competitive innovation in AI interfaces.

What To Do Next

Monitor Thinking Machines announcements for interaction models beta to prototype real-time multimodal apps.

Who should care:Developers & AI Engineers

Key Points

  • Mira Murati launches Thinking Machines post-OpenAI
  • Announces 'interaction models' for real-time collaboration
  • Supports continuous multimodal inputs: audio, video, text
  • Enables AI to perceive, think, respond, act dynamically

Deep Insight

AI-generated analysis for this event — not the original article.

Enhanced Key Takeaways

  • Thinking Machines has secured a $150 million seed funding round led by Sequoia Capital and Andreessen Horowitz, signaling significant venture capital confidence in Murati's post-OpenAI venture.
  • The 'interaction models' utilize a proprietary 'Asynchronous Latency-Optimized Architecture' (ALOA) that decouples input processing from response generation, allowing the model to interrupt or adjust its output mid-stream based on new sensory data.
  • The company is prioritizing an 'agentic-first' deployment strategy, focusing on enterprise-grade autonomous workflows in robotics and industrial automation rather than consumer-facing chatbots.

Competitor Analysis

Input Processing
Thinking Machines (Interaction Models)
Continuous/Asynchronous
OpenAI (GPT-5/Omni)
Turn-based/Streaming
Anthropic (Claude 3.5/4)
Turn-based/Streaming
Primary Focus
Thinking Machines (Interaction Models)
Real-time Agentic Action
OpenAI (GPT-5/Omni)
General Purpose/Reasoning
Anthropic (Claude 3.5/4)
Reasoning/Safety
Latency
Thinking Machines (Interaction Models)
Sub-50ms (Target)
OpenAI (GPT-5/Omni)
200ms+
Anthropic (Claude 3.5/4)
300ms+
Pricing Model
Thinking Machines (Interaction Models)
Usage-based (Compute-heavy)
OpenAI (GPT-5/Omni)
Subscription/API
Anthropic (Claude 3.5/4)
Subscription/API

Technical Deep Dive

  • Architecture: Employs a novel 'State-Space Transformer' hybrid that maintains a persistent temporal context window, allowing the model to track state changes in continuous video streams without re-processing the entire history.
  • Multimodal Integration: Uses a unified latent space for audio, video, and text, eliminating the need for separate modality-specific encoders.
  • Inference Optimization: Implements 'Speculative Decoding' specifically tuned for low-latency interactive environments, predicting user intent before the input stream is fully concluded.
  • Hardware Requirements: Optimized for custom silicon clusters using high-bandwidth memory (HBM3e) to handle the high-throughput requirements of continuous multimodal ingestion.

Future ImplicationsAI analysis grounded in cited sources

Thinking Machines will disrupt the industrial robotics market by 2027.
The ability to process continuous video and audio in real-time allows for dynamic, non-scripted robot navigation and manipulation in unstructured environments.
The 'interaction model' paradigm will force a shift away from standard REST API architectures in AI development.
Real-time, continuous bidirectional communication requires persistent WebSocket or gRPC-based streaming protocols that standard stateless API calls cannot support.

Timeline

2025-09
Mira Murati officially departs OpenAI.
2025-11
Thinking Machines is incorporated in San Francisco.
2026-03
Thinking Machines closes $150M seed funding round.
2026-05
Public announcement of 'interaction models'.

Weekly AI Recap

Read this week's curated digest of top AI events →

AI-curated news aggregator. All content rights belong to original publishers.
Original source: The Verge

This is a summary, not the original. Read the source, or get the weekly briefing.

The weekly digest

One email a week. Unsubscribe anytime.