Mira Murati Presents Interaction-Model Vision for Thinking Machines

At Bloomberg Tech in San Francisco on June 4, Mira Murati, co-founder and CEO of Thinking Machines Lab, made her first major public appearance in roughly 18 months, in an on-stage interview with Bloomberg's Emily Chang, TechCrunch reports. According to TechCrunch, Murati used the conversation to lay out the lab's direction and to discuss its 'interaction models,' the approach Thinking Machines introduced in a research preview in mid-May. TechCrunch writes that the lab has spent about a year and a half raising capital, hiring researchers, and shipping one product, Tinker, an API for fine-tuning open-source models, and that Murati framed the broader product vision as an early step without committing to a release timeline. She also addressed the November 2023 OpenAI board crisis that employees reportedly called the 'blip,' saying she felt clear about her decisions in each moment, per TechCrunch.

Thinking Machines describes interaction models as systems that process audio, video, and text as continuous, parallel streams rather than flattening everything into a single ordered token sequence. In its research-preview write-up, the lab says the model interleaves perceiving roughly 200 milliseconds of input with generating 200 milliseconds of output, letting the system listen, see, and respond at the same time, including interrupting or adding context mid-conversation. Independent coverage of the May preview reported demos such as live two-way translation, repetition counting, and safety monitoring.

Real-time, streaming multimodal interaction is architecturally harder than turn-based, batch prompt flows. Designs of this kind generally require low-latency streaming inference, tightly integrated audio and vision capture, and stateful memory or chunked attention to stay responsive at sub-200-millisecond granularity, which raises compute and engineering costs. The approach also runs directly into established real-time voice and assistant systems, making latency, reliability, and cost per minute the metrics that will decide whether it is practical.

Useful signals include published latency and throughput benchmarks for live streams, demos showing clean interruption and mid-thought handling, developer adoption of Tinker, any open checkpoints, and partner integrations. More detailed disclosures on how the lab manages persistent state, cost, and privacy for streaming audio and video would clarify how close interaction models are to production.