AI Voice Chats Still Feel Awkward because Assistants Don’t Know when to Talk
The startup said the model reached 0.40-second turn-taking latency and outscored Gemini-3.1-flash-live and GPT-realtime-2.0 on real-time benchmarks.
- On Monday, AI startup Thinking Machines announced 'interaction models,' a system achieving 0.40-second turn-taking latency through simultaneous input and output processing. The 'full duplex' architecture mimics the pace of natural human conversation.
- Thinking Machines researchers designed this to solve a 'collaboration bottleneck' where current models freeze perception during generation. The two-part architecture pairs an 'Interaction Model' for constant dialogue with a 'Background Model' for sustained reasoning.
- Native interactivity enables 'backchannel' cues and live translation without interrupting users. In benchmarks like RepCount-A and ProactiveVideoQA, the system successfully engaged with visual data, validating its potential for real-time industrial safety audits.
- While not yet public, Thinking Machines plans to open a 'limited research preview' in the coming months, with a wider release targeted for later this year. The system remains a research project rather than a commercial product.
- Founded last year by OpenAI CTO Mira Murati and former researcher John Schulman, the startup continues expanding its compute ambitions, including a partnership to deploy Vera Rubin systems. This positions the company as a significant contender in real-time AI.
19 Articles
19 Articles
AI that talks back in real time: Mira Murati’s Thinking Machines unveils ‘interaction models’
Thinking Machines Lab, founded by Mira Murati, has introduced "interaction models," an AI approach designed for real-time, natural conversations. This new system continuously processes information and responds simultaneously, unlike current AI that waits for complete prompts.
Mira Murati creating Her with Thinking Machines, year after she broke up with OpenAI and Sam Altman
Mira Murati, former OpenAI CTO who founded Thinking Machines in 2025, is building a new AI system that interacts with users almost like a human would. In demos released on Tuesday, it comes across as something that we saw in the movie Her.
Thinking Machines wants to build an AI that actually listens while it talks
Right now, every AI model you've ever used works the same way. You talk, it listens. It responds, you listen. Thinking Machines is trying to change that by building a model that processes your input and generates a response at the same time, so it's more like a phone call than a text chain.
VentureBeat
technewstube.comThinking Machines shows off preview of near-realtime AI voice and video conversation with new 'interaction models'
By making interactivity native to the model, Thinking Machines believes that scaling a model will now make it both smarter and a more effective collaborator.
Coverage Details
Bias Distribution
- 43% of the sources are Center, 43% of the sources lean Right
Factuality
To view factuality data please Upgrade to Premium
