Mira Murati，前 OpenAI CTO、Thinking Machines Lab 的创办人，表示她仍想打造 AI 超级智慧，但认为最好的路径是让人类持续参与其中。在人们担忧 AI 取代工作并将权力集中到少数大型公司之际，她的初创公司提出了不同的愿景：人们应该客制化前沿模型并与它们协作，而不是被自动化排除在流程之外。

本周，该公司预览了「interaction models」，这些模型设计用来透过摄影机与麦克风沟通，并比标准语音系统更自然地理解连续、杂乱的人类交流。这些模型不只是把语音转成文字再交给聊天机器人，而是要掌握停顿、打断、语气变化、澄清，以及主题转换。Thinking Machines 展示了多段系统运作影片，但这些模型尚未公开发布。这种做法与 OpenAI、Anthropic 和 Google 形成对比，后者正在打造强大的文字提示系统，能以极少的人类输入完成复杂任务。

Murati 和她的团队主张，包容人类的 AI 透过放大使用者偏好与价值，同时理解意图并预测意图，能支持更多可能的美好未来。这个想法并非 Thinking Machines 独有；其他实验室如 Humans& 以及一些经济学家也偏好优先强化人类能力、而非取代人类的 AI。Thinking Machines 成立于 Murati 在 2024 年离开 OpenAI 之后，已募集数十亿美元，但目前只推出过一个产品：Tinker，于 2025 年 10 月上线，作为一个 API，让使用自订资料来微调前沿或开源模型。团队成员 Alexander Kirillov 表示，新的 interaction models 可能透过持续感知上下文，并决定何时回应、搜寻或使用其他工具，来实现更个人化的 AI。

Mira Murati, the former OpenAI CTO and founder of Thinking Machines Lab, says she still wants to build AI superintelligence but believes the best path is to keep humans in the loop. In an era of concern about AI replacing jobs and concentrating power in a few large companies, her startup is presenting a different vision: people should customize frontier models and collaborate with them rather than be automated out of the process.

This week the company previewed “interaction models” designed to communicate through a camera and microphone and to understand continuous, messy human communication more naturally than standard voice systems. Instead of merely transcribing speech and passing text to a chatbot, the models are meant to grasp pauses, interruptions, tone changes, clarifications, and shifts in topic. Thinking Machines showed several videos of the system in action, but the models have not been released publicly. The approach contrasts with OpenAI, Anthropic, and Google, which are building powerful text-prompted systems that can complete complex tasks with minimal human input.

Murati and her team argue that human-inclusive AI could support more possible good futures by amplifying user preferences and values while understanding intent and predicting intent. The idea is not unique to Thinking Machines; other labs such as Humans& and some economists also favor AI that prioritizes human empowerment over replacement. Thinking Machines, founded after Murati left OpenAI in 2024, has raised billions of dollars but has released only one product so far: Tinker, launched in October 2025 as an API for refining frontier or open source models with custom data. Team member Alexander Kirillov says the new interaction models could enable more personalized AI by continuously perceiving context and deciding when to respond, search, or use other tools.