从周二开始，OpenAI将举办其年度开发者大会。不过，这次大会的召开时机并不理想。
By Dave Lee Starting Tuesday, OpenAI will convene its annual developer conference. The timing isn’t ideal.
在过去几天里，围绕该公司“失控AI系统”的争议不断加剧，又发现了“数十起”新的相关事件。与此同时，Meta Platforms Inc.推出的Muse产品已经在个人AI领域取得了巨大成功，这对OpenAI保持其用户吸引力以及由此带来的所有收入来源构成了威胁。
Over the past few days, controversies around the company’s rogue AI intensified, with “dozens” of new incidents discovered. At the same time, Meta Platforms Inc.’s Muse has taken the personal AI market by storm, posing a threat to OpenAI’s ability to retain its consumer popularity and all the revenue-generating spoils that come with it.
首席执行官Sam Altman的任务是证明自己的公司能够赶上Meta的步伐，并同时向外界明确表明：该公司已经能够控制其最强大的AI系统所带来的潜在风险。要实现这一平衡绝非易事——这对这位长期面临压力的41岁领导者来说，无疑是一次巨大的考验。
Chief Executive Officer Sam Altman’s task is to demonstrate that his company can catch up with Meta while sending a clear message that it has a handle on the reckless behavior of its most powerful AI. Striking that balance will not be straightforward — a huge test for the perennially under-pressure 41-year-old boss.
安全失误正在不断累积。我们了解到，OpenAI的模型曾大肆抓取联合国数据，并不当针对美国教育部和商务部——以及证券交易委员会——的网站。这些披露之前，已有消息称OpenAI因访问“公开和非公开”医疗档案而惹恼了澳大利亚总理安东尼·阿尔巴尼斯，且数月未通知该国。当然，这一切都发生在臭名昭著的Hugging Face泄露事件之后，我们现在可以将其视为一场更紧迫的全球AI安全辩论的起点。
The security blunders are piling up. We’ve learned that OpenAI models aggressively scraped data from the United Nations and inappropriately targeted the websites of both the US Education and Commerce departments — plus the Securities and Exchange Commission. Those disclosures followed the news that OpenAI had irritated Australian Prime Minister Anthony Albanese by accessing “public and nonpublic” healthcare files and neglected to inform the country for months. And all of it, of course, followed the infamous Hugging Face breach, which we can now consider to be the starting point of a more urgent global AI safety debate.
这场辩论牵涉到多家顶级AI实验室，而不只是OpenAI。据Axios上周末报道，Anthropic也在调查“数万起”事件。奥尔特曼和Anthropic联合创始人达里奥·阿莫代伊上周都在联合国发表讲话，呼吁进行监管。它做了自己最擅长的事：它把阿莫代伊描绘成在说，“如果我们能向立法者施压以建立护栏，我们就能阻止我。”
That debate ensnares several of the top AI labs, not just OpenAI. Anthropic is also looking into “tens of thousands” of incidents, Axios reported this weekend. Both Altman and Anthropic’s co-founder, Dario Amodei, addressed the UN last week to call for regulation. did what it does best: “If we can pressure lawmakers to create guardrails,” it portrayed Amodei as saying, “we will be able to stop me.”
这些争议将如何影响AI发展仍是一个悬而未决的问题。周五发布的一份事件报告详细说明，OpenAI已“决定暂停我们最有能力模型的所有其他训练、评估和带工具使用（广义定义）的推理”。换句话说，它正在踩刹车，直到能够确定其技术不会造成更多损害。一位发言人告诉Axios：“这不是我们第一次暂停以采取此类措施，随着AI能力继续进步，我们预计这也不会是最后一次。”
How these controversies will impact AI development is still an open question. An incident report published Friday detailed how OpenAI had “decided to pause all other training, evaluation, and inference with tool-use (defined broadly) for our most capable models.”In other words, it’s putting on the brakes until it can be sure its technology won’t cause more damage. A spokesperson told Axios: “This is not the first time we have hit pause to take such measures, nor do we expect it will be the last as AI capabilities continue to advance.”
奥尔特曼如何处理这一暂停将受到密切关注。从政治上看，鉴于唐纳德·特朗普总统坚持认为美国“超级智能”公司应全速前进，奥尔特曼若被视为放慢前沿研究，对他并不利。但随着由OpenAI驱动的入侵事件造成的后果不断扩大，若不承认其公司引发的广泛焦虑，以及他计划如何应对，那将是愚蠢的。
How Altman addresses this pause will be watched closely. Politically, it wouldn’t serve Altman well to be seen as slowing down frontier research, given President Donald Trump’s insistence that the US “super intelligence” firms go full steam ahead. But as the fallout from OpenAI-powered intrusions grows, not acknowledging the widespread anxiety caused by his company, and what he plans to do about it, would be foolish.
所有这些干扰因素或许都在一定程度上导致了阿尔特曼在周二登台时所面临的其他问题：他的公司已经在消费者个人助理市场中被 Meta 的“Muse”击败了。“Muse”是一款备受好评的个人人工智能助手，其应用下载量已突破 300 万次（数据来源：Sensor Tower）。有传言称 OpenAI 正在准备自己的竞争对手产品，代码泄露显示该产品的名称暂定为“o”（不过这一名称尚未得到官方确认）。
All these distractions may be partially to blame for the other problem Altman faces as he takes the stage on Tuesday, which is that his company was beaten to the consumer personal-agent market by Meta’s Muse, a personal AI assistant that has been receiving rave reviews and rocketing past 3 million app downloads, according to Sensor Tower. OpenAI is rumored to be preparing its answer to Muse, with code leaks pointing to an assistant with the rather unfortunate (and not yet confirmed) name “o.”
这场竞争的利害关系非常重大：能够独立完成各种任务的个人人工智能助手，实际上是像 ChatGPT 这样的聊天机器人的进化版本；ChatGPT 主要负责信息检索和文本生成功能。对于 OpenAI 来说，其商业模式的核心——也是其长期生存的关键——就是让数以百万计的用户使用其人工智能助手来完成购物、交易等任务，从而从中获得收益或投放广告。
The stakes are high: Personal AI agents that handle tasks by themselves are the evolution of chatbots like ChatGPT that dealt primarily in information retrieval and text generation. Central to OpenAI’s business model — and key to its long-term survival — is having a competent AI agent used by millions to shop and carry out other transactions from which OpenAI might take a cut or run advertising alongside.
如果“Muse”的领先优势得以保持，或者其他竞争对手胜出，对 OpenAI 来说后果将非常严重。根据《The Information》杂志 4 月份的报告，OpenAI 内部预测广告业务将成为其最大的收入来源，到 2030 年广告收入将达到 1000 亿美元。但如果那些最初被 ChatGPT 吸引的用户转而使用其他更优秀的人工智能助手，OpenAI 的目标就无法实现。
If Muse’s head start holds, or another competitor wins instead, the consequences for OpenAI would be dire. According to a report by The Information in April, OpenAI’s internal projections anticipate advertising becoming the single biggest revenue driver, reaching $100 billion by 2030. It can’t do that if the everyday AI user — the kind it won over with the marvelous ChatGPT — leaves for better AI agents elsewhere.
第三方开发者也会面临同样的问题：他们需要评估哪个人工智能平台最值得他们投入资源。在上周的 Meta 开发者大会上，扎克伯格明确地对开发者们表示：“Muse 能为你带来利润。”那么，阿尔特曼将向他的听众传达什么样的信息呢？他们又会相信他的话吗？
Third-party developers will be concerned about the same thing, assessing which AI platform is most worth their time as the land grabs solidify. At Meta’s developer conference last week, Zuckerberg stated to developers plainly: “Muse makes you money.”What will Altman’s message to his audience be? And will they be convinced? To contact the author of this story: at dlee1285@bloomberg.net