Mira Murati Wants Her AI to 'Keep Humans in the Loop'

"“At some point we will have super-intelligent machines,” Murati tells WIRED. “But we think that the best way to actually have many possible futures-good futures-is to keep humans in the loop.”"

"Murati says AI doesn't need to automate humans out of the equation. A more optimistic approach, she suggests, is to let people build and customize their own frontier AI models, then work with those models to achieve their goals."

"The company's “interaction models” are trained to communicate with a person through a camera and microphone. Unlike many existing voice-mode interfaces, the new models do not simply capture and transcribe speech, then feed it into a language model that processes it in the same way as a chatbot. The interaction models natively understand continuous, messy, human communication-meaning they are better able to grasp the meaning of pauses, interruptions, and changes in tone."

"This allows them to adapt on the fly when someone clarifies a point or changes the subject. The company showed off several videos demonstrating these capabilities, though the models have not been released publicly. Murati's approach stands in contrast to how most big AI companies seem to be pursuing superintelligence today."

Superintelligent machines are expected, but human involvement is presented as essential for creating many good futures. AI is framed as something people can build and customize rather than a system that automates humans out of the equation. Thinking Machines Lab previews “interaction models” trained to communicate with a person using a camera and microphone. These models are designed to understand continuous, messy human communication, including pauses, interruptions, and tone changes. Instead of only capturing and transcribing speech and then treating it like chatbot input, the models adapt in real time when a person clarifies or changes the subject. This approach contrasts with large AI systems that rely on text prompts and require minimal human help.

Read at WIRED