Why OpenAI Is Racing to Make Computer Use Superhuman

在这个版本的 a16z 播客节目中，主持人西玛·安布尔（Sima Amble）和法布里奇奥·塞拉菲尼（Fabrizio Serafini）邀请了来自 OpenAI 的阿里·韦恩斯坦（Ari Weinstein）和詹姆斯·孙（James Sun），共同探讨电脑操作智能体的快速演进、现状与未来。 讨论重点关注了 Astra 的发布，这是 OpenAI 的最新模型，标志着电脑使用能力实现了重大飞跃。虽然先前的模型已经具备很强的能力，但 Astra 带来了关键的增强：它的速度更快，使用的思考标记（thinking tokens）更少，在创意任务（例如在微软画图软件中绘图）上表现更好，并且在边缘情况下的可靠性更高，正朝着 99% 的可靠性迈进。此外，Astra 是 OpenAI 对齐程度最高的模型，具备对提示词注入和非预期操作的强大抵抗力，这为企业部署建立了至关重要的信任。除了实际工作之外，Astra 在玩电子游戏方面也非常精通，它击败了先前的模型，甚至可以连续玩几个小时《辐射》之类的游戏。 对话探讨了面向消费者和企业的不同影响与用例。对于消费者而言，电脑操作智能体充当日常生活中的个人助手——例如 DOTS，这是一个拥有自己的云电脑和网络浏览器的工具，可以处理诸如预约挂号、标记 Gmail 中未确认的日历邀请、主动管理 Slack 消息和 Sentry 日志等繁琐任务。对于企业而言，电脑操作充当了一个通用 API，用于导航缺乏现代 MCP 工具或 API 的传统软件。企业可以跨记录系统（如 ERP、CRM 和财务账目）自动化繁琐的工作流程（如采购、费用对账和市场研究）。 为了确保安全并建立企业信任，OpenAI 引入了确定性控制、沙盒功能以及通过应用程序的“守护者”系统实现的自动批准模式等人在回路（human-in-the-loop）选项。这些工具允许管理员为智能体可以访问的网站或应用程序设置严格的边界，并强制要求对下载、上传或访问新网站等敏感操作进行人工批准。 在谈到智能体如何与应用软件和开放网络交互时，发言者讨论了与网络发布商之间的紧张关系，这些发布商的收入依赖于人类的广告展示，同时也讨论了诸如 WebMCP 等标准的采用，这些标准允许人类和智能体在网站上顺畅协作。 展望未来，OpenAI 的工作重点很大程度上放在提高模型速度上。发言者指出，电脑操作智能体在使用电脑的速度上已经超过了普通人，并且正在迅速接近“超人”般的速度，其表现将优于熟练的电脑用户。这种转变将从根本上改变人们与电脑交互的方式。虽然后台执行仍然是一个主要优势，但实时协作和语音驱动的电脑使用将变得越来越突出。 最后，发言者强调了软件工艺和用户界面的持久价值。尽管廉价软件和自主智能体大量涌现，但人类直觉、直接操作和专门构建的用户界面对于消费者体验的愉悦感和企业的可审计性而言，依然必不可少。初创企业通过将深厚的行业专业知识、市场开拓执行力以及将不同系统连接在一起的垂直软件工作流程结合起来，仍然拥有巨大的机遇。归根结底，行业正在迎来电脑操作智能体的一个重大规模化时刻，架起了强大的后台自动化与连贯的、以人为本的体验之间的桥梁。

In this episode of the a16z podcast, hosts Sima Amble and Fabrizio Serafini are joined by Ari Weinstein and James Sun from OpenAI to discuss the rapid evolution, current state, and future of computer using agents. The discussion highlights the launch of Astra, OpenAI's newest model, which marks a significant leap in computer use capabilities. While previous models were already competent, Astra brings key enhancements: it is faster, uses fewer thinking tokens, performs better on creative tasks (such as drawing in Microsoft Paint), and boasts higher reliability around the edges, moving closer to 99% reliability. Furthermore, Astra is OpenAI’s most aligned model, featuring strong resistance to prompt injection and unintended actions, which builds vital trust for enterprise deployments. Beyond practical work, Astra is exceptionally proficient at playing video games, beating previous models and even playing games like Fallout for hours. The conversation explores the different implications and use cases for consumers and enterprises. For consumers, computer-using agents act as personal helpers in daily life—such as DOTS, a tool that gets its own cloud computer and web browser to handle tedious tasks like booking appointments, flagging unconfirmed calendar invites in Gmail, and proactively managing Slack messages and Sentry logs. For enterprises, computer use acts as a universal API to navigate legacy software that lacks modern MCP tools or APIs. Businesses can automate tedious workflows across systems of record like ERPs, CRMs, and finance ledgers (such as procurement, expense reconciliation, and market research). To ensure safety and build enterprise trust, OpenAI has introduced determinative controls, sandboxing capabilities, and human-in-the-loop options like auto-approve modes via the app's "guardian" system. These tools allow administrators to set strict boundaries on what websites or apps an agent can access and mandate human approval for sensitive actions like downloads, uploads, or visiting new sites. Addressing how agents interact with application software and the open web, the speakers discuss the tension with web publishers whose revenue depends on human ad impressions, as well as the adoption of standards like WebMCP that allow humans and agents to collaborate smoothly on websites. Looking forward, OpenAI is heavily focused on increasing model speed. The speakers note that computer-using agents are already faster than the average human at using a computer, and they are rapidly approaching "superhuman" speed where they outperform expert computer users. This shift will fundamentally change how people interact with computers. While background execution remains a major strength, real-time collaboration and voice-driven computer use will become increasingly prominent. Finally, the speakers emphasize the enduring value of software craftsmanship and user interfaces. Even as cheap software and autonomous agents proliferate, human intuition, direct manipulation, and purpose-built UI will remain essential for both consumer joy and enterprise auditability. Startups still have massive opportunities by combining deep industry expertise, go-to-market execution, and vertical software workflows that bridge disparate systems together. Ultimately, the industry is entering a major scaling moment for computer-using agents, bridging the gap between powerful background automation and coherent, human-centered experiences.