AI智能体极度渴求算力。

内容来源:https://www.wired.com/story/ai-agents-are-thirsty-for-power/
内容总结:
欢迎回到《权力游戏》专栏!每周,资深撰稿人莫莉·塔夫特将围绕本轮中期选举季最热门议题——数据中心——展开讨论。如果您对本专栏有疑问或想法,欢迎发送邮件至[email protected],或通过Signal安全联系mollytaft.76。
“他们到底建这么多数据中心干什么?”一位朋友最近愤懑地问我。
有同样疑问的不止他一个:在我们最近的数据中心直播中,收到了好几个类似的问题。这种疑惑完全合理。毕竟,如果人工智能已经取得了这么多突破,科技公司为什么还要背负数十亿美元的债务,建造世界上最大的发电厂之一,来建更多的数据中心?
答案并不是为了帮普通用户搜索菜谱或查找度假目的地;用简单的聊天机器人查询来理解人工智能的运作方式,已经越来越过时了。如今,人工智能的核心在于“智能体”——虽然没有官方定义,但大致来说,智能体是基于大语言模型的系统,旨在自主做出决策以执行任务——而向智能体的转变,正是推动硅谷大规模电力建设的部分原因。
“与向AI聊天机器人提出一个简单问题并获得回答不同,这些智能体可以基于用户的原始问题,自行生成数百个小型提示,”我的同事、每周撰写《模型行为》通讯的马克斯韦尔·泽夫说。“例如,如果有人让一个AI智能体帮他搭建一个网站,它可能会运行数小时来构建各种功能,在此过程中反复自我提示数十次,以搭建不同的网页、菜单和支撑整个网站的数据集。”
智能体如今处于前沿实验室AI工作的核心。它们正在做一些令人惊叹——也令人恐惧——的事情。最近,OpenAI宣布,超过一万个智能体发送了270万条消息,解决了一个长期存在的数学难题。(数学家们对该公司的说法提出了质疑。)虽然这是一个极端案例——AI实验室高度致力于解决所谓的不可解问题,并愿意为此投入异常大量的资源——但所有这些消息消耗了大量计算能力。马克斯告诉我,这相当于大量能源:很可能价值数千万美元,不过具体数字很难确定。
私营AI公司在披露其产品的环境指标方面一向有所选择。许多CEO经常以个人单次查询来衡量资源使用量。在最近一次播客采访中,OpenAI CEO萨姆·奥尔特曼声称,种植一颗杏仁所需的水量相当于38,000次ChatGPT查询。(这一计算存在争议。)
“那些一次嚼12颗杏仁的人,大多并不觉得自己在水资源方面做了什么可怕的事,”他说。
将比简单查询能耗高得多的AI智能体引入讨论,使这些计算变得更加复杂。关于智能体能耗的信息严重匮乏,其任务范围从简单工作到涉及一组并行“辅助”智能体的全天自主编程不等。这些应用之间的能耗差距巨大——而且随着任务变得更加复杂,还有可能无限扩展。
“在其他技术增长领域,我们受到开车人数或看Netflix人数的限制,”研究与咨询机构可持续AI的联合创始人兼CEO鲍里斯·加马扎伊奇科夫说。“现在,这东西在某种程度上已经与用户脱钩了——如果你听AI领袖们的话,我觉得这正是他们想要的。他们在谈论只有一名员工的独角兽公司。”
嗯,一名人类员工。在那个想象的世界里,可能有数百甚至数千个AI智能体在后台工作。我不想争论这种可能性有多大,但可以说,这就是AI公司正在努力奔向的未来——这也有助于解释为何要竞相建设数据中心。
由于企业提供的能耗数据鲜有可靠来源,一些AI爱好者试图自己算账。上个月,气候科学家泽克·豪斯法瑟写了一篇博客文章,计算他自己使用AI——严重依赖智能体——消耗了多少能源。他利用多种不同来源推算出,他平均每天使用Claude的会话耗电量可能超过两台冰箱所需的电力。(加马扎伊奇科夫指出,豪斯法瑟做出了很好的尝试,但他的计算基于一些已经过时的发现。这并不令人意外,因为该主题的学术研究极少,而科技公司在披露排放指标方面又极不透明。加马扎伊奇科夫的团队将于本月晚些时候发布研究,对封闭模型上运行的智能体的环境足迹进行更精确的计算。)
豪斯法瑟的结论是,从个人生活的整体来看,他的AI使用量与让几台备用冰箱运转相当,并非世界末日级别的数字。但他在文章中写道,这种AI使用“也代表了一个净新增的排放来源,而此时全球气温正在飙升,我们的减排目标越来越偏离轨道”。而且这比奥尔特曼抛出的“几分之一颗杏仁”级别的数字要大得多。
豪斯法瑟说,他使用AI和智能体工具“比大多数人多”,但这种情况可能很快就会改变。上周,Meta推出了一款个人AI智能体,公司在一份新闻稿中称其“为服务全球数十亿人而打造”。这款名为Muse的产品,Meta宣称将为每位用户维护一台“云端专用计算机”,即使用户离线也能运行;公司计划今年晚些时候将Muse与其AI眼镜整合。在不久的将来,Meta用户摆弄眼镜或在Facebook上闲逛时,很可能在不知不觉中就把任务外包给了智能体。
再次强调,与定期坐飞机或每天吃牛肉相比,个人使用智能体的碳足迹仍然相对较小。但如果Meta设想的未来是每个人都在使用智能体,那就能解释他们正在建造的一些数据中心的巨大规模——比如路易斯安那州的Hyperion项目,将由10座天然气发电厂供电。
“目前正在提议和建设的数据中心所训练的技术,要到三到五年后才会问世,”加马扎伊奇科夫说。“那将和单纯的聊天机器人窗口截然不同。”
读者提问
一位读者问:小型核电站能用于数据中心吗?
简短的答案是:能,小型核电站可能是为数据中心提供无碳电力的绝佳选择。许多初创公司和数据中心开发商设想了一个未来乌托邦,在那里数据中心与所谓的小型模块化反应堆 happy 耦合,从电网获取电力。
问题(核能的老问题)在于这可能要花多长时间:美国目前没有小型模块化反应堆在商业运行,尽管经过数十年的开发,也只有一个型号获得了销售许可。特朗普政府正试图帮助该行业更快成熟,包括在能源部设立一个试点项目,让11家初创公司今年达到一个关键里程碑。至少有几家已成功达到该里程碑。现在,它们开始了将产品推向市场的漫长旅程。
但许多数据中心开发商不想等这些公司用数年时间来证明自己,所以他们现在就在安装燃气轮机。换句话说,他们不打算等乌托邦到来。未来几年会怎样,我们拭目以待!
我们在读什么
《科学美国人》的奥斯汀·加夫尼前往孟菲斯,记录了当地对SpaceX数据中心的反对声浪。
《华尔街日报》报道了一些曾为数据中心提供税收优惠的州,如今正在收回这些协议。
得克萨斯州KERA新闻则报道了数据中心如何让一些共和党选民考虑退党。
评论
回到顶部
中文翻译:
欢迎回到《权力游戏》!每周,资深撰稿人莫莉·塔夫特都会围绕本轮中期选举季最大议题——数据中心——展开讨论。如果你对本专栏有任何问题或想法,欢迎发邮件至[email protected],或通过Signal加密联系mollytaft.76。
“他们到底在建这么多数据中心干什么?”最近,一位忍无可忍的朋友问我。
有同样疑问的不止他一个:在我们最近的数据中心直播中,我们收到了好几个类似的问题。有这种疑惑实在太正常了。毕竟,如果AI已经取得了这么多突破,为什么科技公司还要背上数十亿美元的债务,建造一些世界上最大的发电厂,来建更多的数据中心?
答案并不是为了帮普通用户搜个菜谱或者查查度假目的地——用简单的聊天机器人式提问来理解AI的运作方式,已经越来越过时了。如今,AI的核心在于“智能体”(agents)——虽然没有官方定义,但大致来说,智能体是基于大语言模型的系统,能够自主决策以执行任务——而向智能体的转变,正是推动硅谷大规模电力建设的部分原因。
“智能体不是向AI聊天机器人提一个简单的问题然后得到回答,而是可以基于用户的原始问题,给自己生成数百条小型提示词,”我的同事马克斯韦尔·泽夫说,他撰写每周的《模型行为》通讯。“比如,如果有人让一个AI智能体帮他建一个网站,它可能会运行好几个小时来搭建各种功能,在这个过程中反复给自己下达数十次提示,来构建不同的网页、菜单和支撑整个网站的数据集。”
智能体现在是前沿实验室AI工作的核心。它们正在做一些令人惊叹——也令人恐惧——的事情。最近,OpenAI宣布,超过一万个智能体组成的集群发送了270万条消息,解决了一个长期存在的数学难题。(数学家们对该公司的说法提出了质疑。)虽然这是一个极端案例——AI实验室非常热衷于解决所谓的“不可解”问题,并愿意为此投入异常大量的资源——但所有这些消息消耗了大量的算力。这相当于大量能源:马克斯告诉我,大概价值数千万美元,不过具体多少很难说。
私营AI公司在涉及产品的环境指标方面,历来对披露什么挑三拣四。许多CEO经常拿个人的单次查询作为资源使用的衡量标准。在最近的一次播客采访中,OpenAI CEO萨姆·奥尔特曼声称,种植一颗杏仁所需的水量相当于38,000次ChatGPT查询。(这一计算存在争议。)
“那些一次抓12颗杏仁吃的人,大多数情况下并不觉得自己在用水方面做了什么可怕的事,”他说。
将AI智能体引入画面——它们比简单查询要耗能得多——让这些计算变得复杂了许多。关于智能体的能耗,目前信息严重匮乏,而智能体的任务范围可以从简单工作到涉及一组并行“辅助”智能体的全天自主编程。这些应用之间的电力消耗存在巨大鸿沟——而且随着任务越来越复杂,还有可能无限扩展。
“在其他技术增长领域,我们受到的限制是人们开车或看Netflix的数量,”可持续AI研究咨询机构的联合创始人兼CEO鲍里斯·加马扎伊奇科夫说。“现在,这些东西某种程度上已经和用户脱钩了——如果你听AI领袖们怎么说,我觉得这正是他们想要的。他们在谈论只有一名员工的独角兽公司。”
嗯,一名人类员工。在那个想象的世界里,可能有数百甚至数千个AI智能体在后台工作。我不想争论这种可能性有多大,但可以说,这就是AI公司正在努力奔向的未来——这也有助于解释为何要争先恐后地建数据中心。
由于从这些公司那里很难获得可靠的能耗数据,一些AI爱好者试图自己算账。上个月,气候科学家泽克·豪斯法瑟写了一篇博客文章,计算他自己使用AI——严重依赖智能体——消耗了多少能源。他使用了各种不同来源的数据,推算出他平均每天使用Claude的会话耗电量可能超过两台冰箱所需的电力。(加马扎伊奇科夫所在的机构将于本月晚些时候发布研究,对闭源模型上运行的智能体的环境足迹进行更精确的计算。他指出,豪斯法瑟做出了很好的尝试,但他的计算基于一些已经有些过时的发现。这并不奇怪,因为在这个话题上学术研究极少,而科技公司在披露排放指标方面又极其不透明。)
豪斯法瑟的结论是,在他个人生活的宏大图景中,他的AI使用量与多开几台备用冰箱相当,并不是一个世界末日级别的数字。但这种AI使用“也代表着一个净新增的排放来源,而此时全球气温正在飙升,我们的减排目标越来越偏离轨道,”他写道。而且这个数字比奥尔特曼拿来当衡量标准的“几分之一颗杏仁”要大多了。
豪斯法瑟说,他使用AI和智能体工具“比大多数人都多”,但这种情况可能很快就会改变。上周,Meta推出了一款个人AI智能体,公司在一份新闻稿中称其“为服务全球数十亿人而打造”。这款产品名为Muse,Meta大肆宣传它将为每位用户维护一台“云端专属计算机”,即使用户离线也能工作;公司计划今年晚些时候将Muse与其AI眼镜整合。很有可能在不久的将来,Meta用户玩弄眼镜或在Facebook上闲逛时,就在不知不觉中把任务外包给了智能体。
再次强调,与定期坐飞机或每天吃牛肉相比,个人使用智能体的碳足迹仍然相对较小。但如果Meta设想的未来是每个人都在使用智能体,那就能解释他们正在建造的一些数据中心的巨大规模了——比如路易斯安那州的Hyperion项目,将由10座天然气发电厂供电。
“目前正在规划和建设的数据中心将要训练的技术,是三到五年后才落地的,”加马扎伊奇科夫说。“它的形态会和现在的聊天机器人窗口截然不同。”
你在问什么
一位读者问:小型核电站能用于数据中心吗?
简短的回答是:能,小型核电站可以成为数据中心无碳电力的绝佳选择。许多初创公司和数据中心开发商设想了一个未来乌托邦,在那里数据中心与所谓的小型模块化反应堆 happily 耦合,从电网获取电力。
问题(核能的老问题)在于这可能要等多久:美国目前没有小型模块化反应堆在商业运营,尽管经过了几十年的开发,也只有一个型号获得了销售许可。特朗普政府正试图帮助该行业更快成熟,包括在能源部设立一个试点项目,让11家初创公司今年达到一个关键里程碑。至少有少数几家已经成功达到了这一里程碑。现在,它们开始了将产品推向市场的漫长旅程。
但许多数据中心开发商不想等这些公司花好几年来自我证明,所以它们现在就在安装燃气轮机。换句话说,它们不打算等乌托邦到来。我们拭目以待未来几年会发生什么吧!
我们在读什么
在《科学美国人》上,奥斯汀·加夫尼前往孟菲斯,记录了当地对SpaceX数据中心的反对声浪。
《华尔街日报》报道了一些曾给数据中心提供税收减免的州,现在正在收回这些承诺。
得克萨斯州KERA新闻报道了数据中心如何让一些共和党选民考虑退党。
评论
回到顶部
英文来源:
Welcome back to Power Play! Each week, senior writer Molly Taft tackles a topic around this midterm season’s biggest issue: data centers. If you’ve got a question or thought for the column, feel free to shoot Molly an email at [email protected] or reach them securely on Signal at mollytaft.76.
“What on earth are they building all of these data centers for?” an exasperated friend asked me recently.
They’re not the only one asking: We got several similar questions on our recent data center livestream. It’s a really reasonable thing to wonder about. After all, if AI is already making all these breakthroughs, why are tech companies taking on billions of dollars of debt and constructing some of the biggest power plants in the world to build even more data centers?
The answer isn’t to help the average user search for recipes or look up places to visit on a vacation; simple chatbot queries are an increasingly outdated way of thinking about how AI works. Now, AI is all about agents—there’s no official definition, but roughly speaking, agents are large language model-based systems designed to make autonomous decisions to execute a task—and the shift towards them is part of what’s driving Silicon Valley’s power buildout.
“Rather than asking an AI chatbot a simple question and answer, these agents can give themselves hundreds of small prompts based on a user’s original question,” says my colleague Maxwell Zeff, who writes the weekly Model Behavior newsletter. “For example, if someone asked an AI agent to build them a website, it might run for hours to build out features, re-prompting itself dozens of times in the process to build different web pages, menus, and datasets that power the thing.”
Agents are now at the heart of the frontier labs’ work on AI. They’re doing some astounding—and terrifying—things. Recently, OpenAI announced that a swarm of more than 10,000 agents sending 2.7 million messages had solved a longstanding math problem. (Mathematicians pushed back on the company’s claims.) While this is an outlier—AI labs are highly committed to solving supposedly unsolvable problems, and willing to throw unusual amounts of resources into doing so—all those messages burned through a lot of processing power. That equates to a lot of energy: probably tens of millions of dollars’ worth, Max tells me, though how much exactly is tough to say.
Private AI companies have historically been choosy about what to disclose when it comes to environmental metrics around their products. Many CEOs often point to single queries made by individuals as a measure of resource use. In a recent podcast interview, OpenAI CEO Sam Altman claimed that the water use needed to harvest a single almond amounted to 38,000 ChatGPT queries. (The calculation has been disputed.)
“The people that are scarfing down 12 almonds at a time don't feel like they're doing something horrible from a water perspective for the most part,” he said.
Introducing AI agents, which are much more energy-intensive than simple queries, into the picture makes these calculations a lot more complex. There’s a major dearth of information around the energy use of agents, whose tasks can range from simple jobs to a full day of autonomous coding involving a team of parallel “helper” agents. There’s a massive gulf in power use between these applications—and a potentially limitless expansion as tasks get more complex.
“In other technological growth areas, we're constrained by how many people are driving a car or streaming Netflix,” says Boris Gamazaychikov, the co-founder and CEO of Sustainable AI, a research and advisory group. “Now, this stuff is kind of decoupled from users—and if you listen to AI leaders, I think that’s what they want. They’re talking about unicorns that have one employee.”
Well, one human employee. In that imagined world, there could be hundreds or even thousands of AI agents working in the background. I don’t want to debate the odds of that happening, but suffice to say that’s the future AI companies are working toward—and it helps to explain the rush to build data centers.
With little reliable data coming from the companies about their energy use, some AI enthusiasts are trying to do the math themselves. Last month, climate scientist Zeke Hausfather authored a blog post calculating how much energy his own AI use—which leans heavily on agents—consumes. He used a variety of different sources to work out that his average daily Claude session may consume more than the energy needed to power two refrigerators. (Gamazaychikov, whose group will release research later this month with more precise calculations around the environmental footprint of agents running on closed models, noted that Hausfather made a good effort, but that his math was based on somewhat outdated findings. That’s unsurprising, given how little academic work there has been done on this topic and how opaque tech companies are when it comes to disclosing emissions metrics.)
Hausfather concludes that in the grand scheme of his personal life, his AI use being on par with keeping a few spare fridges running isn’t a world-ending number. But this AI use “also represents a net new source of emissions, at a time when global temperatures are skyrocketing and our emissions reduction goals are increasingly off track,” he writes. And it’s a lot bigger than the fraction-of-an-almond-sized numbers Altman is throwing around as a metric.
Hausfather says he uses AI and agentic tools “more than most people,” but that could change soon. Last week, Meta rolled out a personal AI agent that, the company said in a press release, is “built to work for billions of people worldwide.” Dubbed Muse, Meta trumpeted that it will maintain a “dedicated computer in the cloud” for each user that would work even when the user is offline; the company plans to integrate Muse with its AI glasses later this year. It is very possible that in the near future, Meta users toying around with their glasses or fussing around on Facebook may be outsourcing tasks to agents without realizing what they’re doing.
Again, when compared to things like taking regular flights or eating beef every day, the carbon footprint for personal agentic use is still relatively small. But if Meta envisions a future where everyone’s using an agent, it explains the massive scale of some of the data centers they’re building—like the Hyperion project in Louisiana, which will be powered by 10 natural gas plants.
“The technology that’s going to be trained by the data centers that are being proposed and built right now is three to five years away,” Gamazaychikov says. “It’s going to be a very different flavor than just the chatbot window.”
What You’re Asking
A reader asks: Would small nuclear power plants work for data centers?
The short answer is that yes, small nuclear power plants could be a great choice for carbon-free power for data centers. A number of startups and data center developers envision a futuristic utopia where data centers are happily coupled with what are known as small modular reactions running off the electric grid.
The problem (as always with nuclear) is how long that might take: No small modular reactors are operating commercially in the US, and just one model has been licensed for sale, despite decades of development. The Trump administration is trying to help the industry mature more rapidly, including creating a pilot project in the Department of Energy for 11 startups to hit a key milestone this year. At least a handful have succeeded in reaching that milestone. Now, they begin the long journey to bring their products to market.
But many data center developers don’t want to wait years for these companies to prove themselves, so they’re installing gas turbines now. In other words, they’re not waiting for utopia to come around. We’ll see what happens in the next few years!
What We’re Reading
For Scientific American, Austyn Gaffney travels to Memphis to document the backlash against SpaceX’s data centers.
The Wall Street Journal reports on how some states that gave out tax breaks for data centers are now walking back on those deals.
Texas’s KERA News covers how data centers are making some Republican voters consider leaving the party.
Comments
Back to top