Thinking Machines builds an AI that listens while it talks
▼ Summary
– Thinking Machines Lab announced interaction models that allow AI to interrupt users, unlike current models that follow a strict input-output sequence.
– The model uses “full duplex” technology to process input and generate responses simultaneously, mimicking a phone call rather than a text chain.
– The company claims TML-Interaction-Small responds in 0.40 seconds, matching natural conversation speed and outperforming comparable models from OpenAI and Google.
– This is a research preview, not a product, with a limited release in the coming months and wider release later this year.
– The benchmarks are impressive and the native interactivity idea is interesting, but real-world performance remains unverified until public testing.
Thinking Machines Lab, the AI startup launched last year by former OpenAI CTO Mira Murati, unveiled a new concept on Monday that it calls interaction models. At its core, this technology enables AI to interrupt you during a conversation.
Every AI model currently available operates on the same principle: you speak, it processes, then it replies. Thinking Machines aims to rewrite that script. By building a model that can process input and generate output simultaneously, the company hopes to make interactions feel more like a natural phone call than a stilted exchange of text messages.
The technical name for this capability is “full duplex” communication. According to the company, its model, TML-Interaction-Small, achieves a response time of 0.40 seconds. That speed matches the pace of natural human dialogue and outpaces comparable offerings from industry giants like OpenAI and Google.
Before anyone gets too excited, this remains a research preview, not a finished product. There is no public release yet. Thinking Machines plans to offer a “limited research preview” in the coming months, with a broader rollout scheduled for later this year.
So what should we make of this? It is still hard to say. The benchmarks are undeniably strong, and the core concept is compelling: making interactivity a native feature of the model rather than a clumsy add-on. Whether the real-world performance will match the technical claims, however, remains an open question. That answer will only come when users can finally test it themselves.
(Source: TechCrunch)
