字号·· | 护眼
彭博社

AI代理是寻找问题的解决方案AI Agents Are a Solution Looking for a Problem

点「原文对照」整页切到原文,或双击某段只看那段的原文。

面向企业的AI智能体一直难以获得市场青睐。摄影师:Joan Cros/NurPhoto,图片来源:Getty ImagesAI agents for business have been slow to gain traction Photographer: Joan Cros/NurPhoto via Getty Images

向能够执行任务而非仅生成内容的人工智能模型转型,尚未以科技行业迫切需要的方式起飞。

The shift to artificial intelligence models that can carry out tasks, not just generate content, hasn’t yet taken off in the way the tech industry desperately needs it to.

Salesforce Inc. 旨在为企业服务的 Agentforce 服务推出近两年来,一直难以获得市场认可。澳大利亚联邦银行在推出智能体后,用户来电激增,被迫撤回了削减 45 名客服人员的计划。Gartner 的一项评估发现,企业部署智能体的项目中,不足 10% 能带来“强劲”的投资回报。有人可能会说,我们仍处于这项技术的试验阶段,但考虑到涉及的巨额资金,这种状况不可能再维持两三年。

Salesforce Inc. has struggled to get traction with its nearly two-year-old Agentforce service for businesses. Australia’s Commonwealth Bank reversed plans to cut 45 customer-service staff after its rollout of agents led to a surge in user calls. And a Gartner review found less than 10% of agent deployments by businesses were leading to “robust” return on investment. You could argue that we’re still in the experimentation phase for this technology, but it can’t remain this way for another two or three years with so much money on the line.

数万亿美元的数据中心支出,很大程度上建立在这样一个设想之上:智能体而非人类,将消耗大部分新增算力,为企业和消费者执行无数任务。要证明这类投资的合理性,AI 智能体的使用量需要增长约 2000% 至 3000%。

The trillions of dollars of spending on data centers is largely underpinned by the idea that agents, rather than humans, will burn through most of the new computing power to carry out myriad tasks for businesses and consumers. To justify that kind of investment, the use of AI agents would need to increase by around 2,000% to 3,000%.

目前,大众似乎对智能体持观望态度。Uber Technologies Inc. 首席财务官 Balaji Krishnamurthy 本周在财报电话会上表示,在万圣节那一周,Uber 和 Uber Eats 处理的 3.2 亿次出行中,仅有 500 单外卖配送和 4 单网约车是通过智能体完成的。“目前我们看到的仍处于非常非常早期,且增长极其有限,”他补充道。

So far, the general public seems wary of agents. Out of the 320 million trips processed by Uber Technologies Inc. and Uber Eats over the week of Halloween, just 500 food deliveries and four ride-hailing trips were made via agent, the company’s Chief Financial Officer Balaji Krishnamurthy said this week on an earnings call. “It is very, very early and very limited traction that we have seen so far,” he added.

Meta Platform Inc. 的 Muse 是为数不多较为成功的智能体之一,在发布约六周内,其在美国和加拿大的用户数达到 1000 万,其中超过四分之一的人每天都在使用,常用于帮助寻找网购优惠。(Muse 尚未在欧洲推出。)据市场情报公司 Sensor Tower 数据,其留存率在一周后为 30%,优于 ChatGPT。

Meta Platform Inc.’s Muse as been one of the more successful agents, hitting 10 million users in the US and Canada within about six weeks of its launch, with more than a quarter of those people using it daily, often to help them shop for online deals. (Muse isn’t available in Europe yet.) It also has a retention rate of 30% after one week, better than that of ChatGPT, according to market intelligence firm Sensor Tower.

但 Muse 也存在问题。它无法访问 Apple Inc.、Amazon.com Inc. 或几家大型银行的服务,一些用户发现自己不得不“照看”该代理——核实它是否正确读取了账单,或在购买前确认购物车中的商品——这在一定程度上是因为它会产生幻觉。

But Muse has problems. It can’t access services from Apple Inc., Amazon.com Inc. or several large banks, and some users have found themselves babysitting the agent — verifying that it read a bill properly or confirming items in a cart before purchase — not least because it can hallucinate.

一名叫 Ugo Nwune 的软件工程师最近在 Threads 上抱怨,Muse 给他发了一条消息,说在他的 Gmail 账户中找不到来自 Anthropic PBC 的邮件,尽管他并未要求它这样做。Nwune 觉得这件事令人不安,于是停止使用 Muse。Meta 后来表示,是一个 Bug 导致模型产生幻觉,该问题已得到修复。

A software engineer named Ugo Nwune recently complained on Threads that Muse had sent him a message about how it couldn’t find an email from Anthropic PBC in his Gmail account, even though he hadn’t asked it to do so. Nwune found the incident disconcerting enough to stop using Muse. Meta later said that a bug caused the model to hallucinate and that the issue had been fixed.

企业面临同样的信任挑战。生成式 AI 的承诺在于,它能提供通用辅助并处理不可预测的问题。但 Gartner 发现,AI 代理在针对非常具体的任务进行设计时效果最好——比如将发票与采购订单匹配,或处理薪资问题——而不是用于广泛的自主工作。

Businesses face the same trust challenge. The promise of generative AI is that it can offer general-purpose assistance and deal with unpredictable problems. But Gartner has found that AI agents work best when designed for very specific tasks — like matching invoices with purchase orders or handling payroll questions — rather than for broadly autonomous work.

如果代理在遵循明确规则时最有效,这就引出了一个令人不安的问题:更便宜的传统软件难道不能做到这一点吗?行业对扩展 AI 能力的痴迷,在某些情况下意味着客户正在不必要地用昂贵的新技术取代旧的自动化手段。代理可能仍是一把在寻找钉子的锤子。

If agents are most effective when they follow clear rules, that begs an uncomfortable question: Can’t cheaper, traditional software do that? The industry’s obsession with expanding AI’s capabilities has, in some cases, meant its customers are needlessly replacing old automation with expensive new tech. Agents may still be a hammer looking for a nail.

这呼应了 AI 商业中一个更广泛、长期存在的问题。打造像电力一样具有革命性的通用技术,可能会造成决策瘫痪,让最终用户不确定在何处以及如何使用它。科技公司应更多地专注于为企业客户提供指导,说明代理能在哪些方面解决实际问题。

That echoes a broader, perennial problem with the business of AI. Making general-purpose technology that’s as revolutionary as electricity can create decision paralysis and leave end users unsure of where and how to use it. Tech firms should focus more on giving their enterprise customers guidance on where agents can solve real problems.

面向消费者的智能体已在一定程度上实现了这一点。Meta 的 Muse 设有“创意”标签页,可根据用户过往的对话和活动,或用户分享的更宏观目标,推荐其可提供协助的方式。总部位于旧金山的 Wabi 是另一款面向消费者的 AI 智能体,它为执行的每项任务创建带有图片或地图的独立迷你应用,同时邀请同样使用该应用的好友给智能体发消息,并就这些任务相互交流。

Consumer-focused agents already do this to some extent. Meta’s Muse has an “Ideas” tab that suggests ways it can assist you based on your past conversations and activity or broader goals you’ve shared. San Francisco-based Wabi, another AI agent for consumers, creates separate mini-apps with images or maps for each task that it carries out, while inviting friends who are also using the app to message the agent and each other about those tasks.

该应用可能会创建一个文本对话串,让好友们投票决定周末晚餐的聚餐地点,并分享地图和建议行程。Wabi 创始人尤金妮亚·库伊达表示,智能体应该“不仅仅是文本”,而应包含视觉元素和应用,以用户已熟悉的方式(如群聊)与用户沟通。

The app might create a text thread where friends can vote on where to meet for a weekend dinner, sharing a map and suggested itinerary. Agents should be “more than just text” and include visuals and apps, communicating with users in ways that are already familiar, like a group chat, says Wabi’s founder, Eugenia Kuyda.

她已将一个智能体变成了自己的通讯和行政中心,与家庭成员分别建立文本对话串来安排活动。例如,Wabi 每天会向库伊达和她的男友提出一个关于他们关系的问题,只有在双方都作答后才会揭示对方的回答。

She’s turned an agent into a communications and admin hub for herself, creating separate text threads with her family members to organize activities. For instance, Wabi asks Kuyda and her boyfriend a question about their relationship each day, and only reveals their answers when they’ve both replied.

构建智能体的人或许需要更具创造性地思考,以提供类似这样新颖独特的功能。试想社交媒体网站如何通过点赞和评论引入了新的互动方式。智能体同样可能开创我们未曾设想的实用工作或沟通方式。

The people building agents probably need to think more creatively to offer similarly fresh and unusual functions. Consider how social media sites introduced new ways of interacting through likes and comments. Agents, too, could usher in useful ways of working or communicating that we haven’t considered before.

但从目前人们和企业使用智能体的方式来看,其未来恐比行业那些宏大(且昂贵)的预测所勾勒的图景更为温和。成功的产品或许不会接管我们生活的大片领域,而只是让我们能将狭窄、特定类型的繁琐事务交由其处理。这意味着,智能体距离引发足以证明数万亿美元 AI 基建投入合理的需求爆发,可能还有很长一段路要走。科技行业必须为其日益昂贵的锤子找到更多钉子。

But the way people and businesses are using agents today points to a more modest future than the one sketched out by the industry’s grand (and expensive) predictions. The winning products might not take over broad swathes of our lives, but rather allow us to hand off narrow, specific kinds of drudgery. That means agents may still be a long way from the demand explosion needed to justify a multi-trillion-dollar AI buildout. The tech industry must find more nails for its increasingly expensive hammer.

阅读媒体原文 ↗