Mira Murati's Thinking Machines demos real-time 'interaction models'

Thinking Machines, the AI company founded by former OpenAI CTO Mira Murati, announced Monday it's developing "interaction models" that continuously process audio, video, and text while responding in real time.

The company describes current AI models as operating in a "single thread" where they must wait for complete user input before responding, with perception frozen during generation. According to Thinking Machines, this creates a "bandwidth bottleneck" that limits collaboration between humans and AI.

Interaction models aim to solve this by processing multiple modalities simultaneously. The company demonstrated examples including listening for animal mentions in stories, real-time speech translation, and posture detection alerting users when they slouch.

Technical approach

Thinking Machines frames the problem as fundamentally architectural. Current models experience "no perception of what the user is doing or how the user is doing it" until input is complete, and "perception freezes" during generation until the model finishes or is interrupted.

The company's approach enables models to "meet humans where they are, rather than forcing humans to contort themselves to AI interfaces," according to its announcement.

Availability

No technical specifications, benchmark scores, or pricing have been disclosed. Thinking Machines plans to open a "limited research preview" in the "coming months" with a wider release later in 2026. The models are not currently available for testing.

Company background

Murati founded Thinking Machines in February 2025 after departing OpenAI, where she served as CTO. The startup has experienced significant staff turnover, with key members leaving for Meta and returning to OpenAI.

What this means

Thinking Machines is attempting to solve a genuine limitation in current LLM architecture: the turn-based nature of inference. If successful, simultaneous multimodal processing could enable more natural human-AI interaction, particularly for collaborative tasks. However, without technical details on latency, compute requirements, or quality degradation during simultaneous processing, it's unclear whether the approach represents a fundamental architectural advance or an engineering optimization. The company's staff retention issues and lack of concrete deployment timeline suggest early-stage development.