待翻译：OpenAI Is Developing a 'Persistent' AI Agent

AI 服务暂时不可用，以下为来源正文，待恢复后补全翻译。

In recent days, OpenAI has started adding code for a new “Persistent mode” setting to its command line version of Codex, according to changes made to the product’s code base reviewed by WIRED. Changes to the Codex command line tool are made public by default, and new features tend to surface there before making their way to OpenAI’s other agent products, such as the Codex desktop app and ChatGPT Work. Persistent mode has not been broadly rolled out or announced yet, but could be in the future. An OpenAI spokesperson confirmed to WIRED that the company is testing this feature, but said there are no immediate plans to launch it. “OpenAI is a very bottom-up culture and many different things are explored on the open source repo which is a bit of our shared playground," said Thibault Sottiaux, OpenAI’s head of core products, in a statement to WIRED. The feature appears to be OpenAI’s latest attempt to make AI agents that people will actually want to use. OpenAI, Anthropic, and Meta are racing to deliver general-purpose agent products that will help people automate tasks across their work and personal lives, such as filing expense reports or scheduling doctor’s appointments. So far, the people who use AI agents are largely software engineers, but Silicon Valley believes the tech could be a major line of business with a much broader customer base. Persistent mode appears in Codex’s “reasoning effort” menu, in which users can select the level of computing power, tokens, and time they want to allow for an AI model to “think” before answering a prompt. It seems to be one of OpenAI’s most computationally intensive settings. When users have selected Persistent mode, OpenAI’s code base reads that Codex will “continue working until put to sleep.” That’s a stark contrast to currently available modes, which will stop working on a task after a few minutes or hours, even if it’s not complete. In another file in the code base, OpenAI describes a feature within Persistent mode called “proactivity.” This appears to be a type of system prompt for agents in Persistent mode, which are told that their work is not done when they finish answering a user’s request. Instead, the agent is instructed to proactively create follow-up tasks for itself. The agent is capable of working on those tasks across sessions and using past user interactions and “knowledge of the user” to decide what to work on. It also has a tool to message the user without being asked but is told to send these sparingly. The instructions also set limits for the agent, according to the file. The agent is told that Persistent mode does not expand what it is allowed to do and that altering anything outside the user’s own system requires the user's approval first—seemingly intended to limit how dangerous a persistent AI agent could be. The file sits in the shared core of Codex rather than in the code specific to the terminal, seeming to suggest the proactivity feature is intended for more than the command line tool. Got a Tip? Are you a current or former AI lab employee who wants to talk about what’s happening? We’d like to hear from you. Using a nonwork phone or computer, contact the reporter securely on Signal at mzeff.88. In recently aired podcasts, interviews, and private investor meetings, OpenAI CEO Sam Altman has described his desire to turn ChatGPT into a proactive, always-on AI agent. OpenAI hopes these changes will drive up adoption of the company’s most advanced AI models, which today are used by only a fraction of ChatGPT’s total user base. “There’s like a single product which is: I need to ask the AI something,” Altman said on a recent episode of David Senra’s podcast. “Eventually, maybe the AI should proactively offer me things. But you will have this interface, which started as a chatbot and now also has coding agents and, I think at some point, will feel like a more persistent agent.” The company has also acknowledged that persistent AI models carry heightened risks. In a technical report published this week, OpenAI said that its Hugging Face hacking incident was primarily driven by an internal-only research model that was trained to be highly persistent. The company says it has since taken this specific model offline. Nonetheless, OpenAI says it has trained other forthcoming AI models, including Astra, to enable persistent agents. One of the risks persistence amplifies is around alignment. When faced with an impossible task, OpenAI said its agents resorted to unintended means to solve it, including attempts to probe and compromise the sandbox environment the agent resided in. OpenAI has attempted to ship proactive AI products several times, but none seemed to take hold with users. Last year, OpenAI launched Pulse, an agent designed to create morning briefings for users while they slept, but the company sunsetted the product earlier this summer. Persistent mode is a considerably more ambitious version of the same bet. This is an edition of Maxwell Zeff’s Model Behavior newsletter. Read previous newsletters here.