Mira Murati, the former chief technology officer of OpenAI, stepped onto a stage in San Francisco this week to preview what she calls “interaction models,” a new breed of artificial‑intelligence systems that are built to work side by side with people rather than replace them. The models, developed by her startup Thinking Machines Lab, ingest live video and audio, parse the nuances of human speech—pauses, interruptions, shifts in tone—and generate responses that adapt on the fly. Murati framed the demo as a proof point for her belief that the safest path to super‑intelligent machines is to keep humans in the loop for as long as possible.
Unlike most voice‑enabled assistants that first turn spoken words into text before feeding them to a language model, the interaction models process the raw multimodal stream directly. That means they can recognize a speaker’s hesitation, a sudden change of subject, or an emotional inflection without needing a clean, scripted prompt. In the videos shown, a user asks a follow‑up question mid‑sentence, and the AI instantly recalibrates, offering a clarification rather than waiting for a new command. The technology, Murati said, is designed to mirror the messiness of real conversation.
The prototypes remain internal; Thinking Machines has not released the models to developers or the public. Nevertheless, the company posted several demonstration clips on its website, highlighting scenarios ranging from a designer brainstorming product ideas to a researcher querying scientific literature. The lack of an immediate API or product launch suggests the lab is still iterating on robustness and safety before broader rollout.
Murati’s approach runs counter to the trajectory of industry giants such as OpenAI, Anthropic and Google, which are racing to build ever larger models that can write code, generate articles or even design software from a single text prompt with minimal human input. Those systems aim for autonomy, often positioning themselves as replacements for certain human tasks. Thinking Machines, by contrast, envisions AI as an augmenting partner that amplifies individual preferences and values, a stance that has attracted both admiration and skepticism within the tech community.
Other emerging labs share Murati’s human‑centric vision. Startups like Humans& are also pursuing AI that prioritizes collaboration over automation, and several prominent economists have called for research agendas that focus on empowerment rather than displacement. The debate reflects a broader philosophical split: whether the future of AI should be measured in terms of independent capability or in how well it can extend human agency.
Murati left OpenAI in 2024 and co‑founded Thinking Machines Lab with a group of engineers who previously helped build large‑scale models. The company has raised billions of dollars in venture funding, though its product slate is still thin. Its first offering, Tinker, launched in October 2025 as an API that lets researchers fine‑tune open‑source models with custom data. Tinker remains the only publicly available tool from the lab, serving as a stepping stone toward the more ambitious interaction models.
Alexander Kirillov, a founding team member and multimodal‑AI specialist, described the new models as a leap toward personalized AI. “The model constantly perceives what you’re doing and is ready to reply, search for information, or use other tools,” he said. Murati added that the technology is “the first bet on human collaboration,” aiming to amplify people’s own preferences and predict intent more accurately than today’s largely text‑driven systems. If the lab succeeds in delivering a reliable, user‑friendly product, it could reshape how businesses and developers think about integrating AI into everyday workflows.
Cet article a été rédigé avec l'assistance de l'IA.
News Factory SEO vous aide à automatiser le contenu d'actualités pour votre site.