当我向我母亲介绍 Uber 时，她问这是不是总是同一个司机。想象一下！那可怜的家伙整天等着，生怕我需要用车。显然，Uber 的创新在于把我与附近任何可用的司机配对。只有这样才经济——而且这个想法花了十多年才实现盈利。
When I introduced my mother to Uber, she asked whether it was always the same driver. Imagine that! Some poor guy waiting around all day on the off chance I’d need a car. Uber’s innovation, obviously, was to pair me with any available driver nearby. Only then was it economical — and it still took more than a decade for the idea to turn a profit.
我的意思是，只要不考虑成本，打造受欢迎且有用的服务其实很容易。想想那首关于可疑商业模式的好歌：《The drinks were free》。Meta Platforms Inc. 的 AI 助手 Muse 正处于这种状态：虽然是一次突破性的成功，但首席执行官马克·扎克伯格在周三承认，它太昂贵，无法大规模推广，因为他基本上是免费提供的。
My point is, it’s easy to build a popular and useful service if money is no object. Think , a great song about a questionable business model: The drinks were free. That’s where Meta Platforms Inc.’s AI assistant, Muse, is today: a breakout success, but one that Chief Executive Officer Mark Zuckerberg admitted on Wednesday is too expensive to scale because he’s basically giving it away.
扎克伯格在公司第三季度财报电话会议上对投资者说：“迄今为止最大的限制之一是，这是一项计算密集型服务，我们尚未在全球范围内推出。”如果 Meta 已经如此推出，当前用户数量——超过 1000 万，其中 25% 每天使用——可能会大幅增加。但扎克伯格补充说：“我们实际上必须去做工作，使效率提升到合理的水平，才能让这件事有意义。”
“One of the biggest constraints so far is that because this is a compute-intensive service to run, we have not launched it globally yet,” Zuckerberg told investors on the company’s third-quarter earnings call. Had Meta done so, the current number of users — more than 10 million, with 25% of them using it every day — would likely be considerably larger. But “we actually have to go do the work to bring the efficiency in line to make that make sense,” Zuckerberg added.
如果扎克伯格要实现其既定目标——让 Muse 用户超过 10 亿，且不让公司破产，挑战是巨大的。Muse 之所以取得进展，是因为它在执行一系列任务方面很有能力，但许多任务的完成效率很低。上周我让它帮我在纽约市找一场伊丽莎白·霍尔姆斯纪录片的放映，它的做法是调用 Meta 为我在某数据中心预留的“虚拟机”上的浏览器，开始搜刮电影院网站。
The challenge is significant if Zuckerberg is to reach his stated goal of more than a billion Muse users without bankrupting the company. Muse is making inroads because of how capable it is at performing a range of tasks, but many of these are completed inefficiently. When I asked it to find me a showing of the Elizabeth Holmes documentary in New York City this past week, its process was to load up a browser on the “virtual machine” Meta has reserved for me in a data center somewhere and start combing movie theaters’ websites.
每位 Muse 用户都能获得这样一台虚拟机。从某种意义上说，Meta 迄今已赠送了 1000 万台 PC。与仅在你提问时才运行的聊天机器人不同，AI 代理旨在几乎持续不断地在后台为你工作，甚至在你睡觉时也是如此。Muse 的每一个动作都意味着 Token、Token、Token（钱、钱、钱）。
Every Muse user gets one of these virtual machines. Meta has so far given away 10 million PCs, in a sense. Unlike chatbots, which run only when you ask them questions, AI agents are meant to be almost constantly working for you in the background, even when you sleep. Everything Muse does means tokens, tokens, tokens (money, money, money).
底线是，扎克伯格估计，算力成本必须至少下降十倍，账目才能算得过来。即便如此，Muse 仍指望其众多创收构想——订阅费、应用内交易抽成、企业服务费、广告——能够兑现，Muse 才有可能接近盈亏平衡。
The bottom line is that Zuckerberg estimates computing costs will need to fall at least tenfold before the numbers start to work. And even then, Muse is relying on its many revenue-generating ideas — subscriptions, cuts of transactions made through the app, fees from businesses, ads — to come through before Muse can get close to turning a profit.
好消息是，从大体上讲，根据一项研究，要达到相似的性能基准，运行 AI 模型的成本确实每年下降约 5 到 10 倍。然而，随着这些助手开始变得更受欢迎，竞争开始升温，Muse 的能力不能停滞不前。那就像苹果试图年复一年发布完全相同的手机。根据 Artificial Analysis 的数据，Muse 目前运行在 Muse Spark 上，这是一个每项任务成本为 1.37 美元的模型。
The good news is that, broadly speaking, the cost of running AI models does indeed fall about five to 10 times each year to achieve similar performance benchmarks, according to one study. However, as these assistants start to become more popular, and competition starts heating up, Muse’s capabilities can’t stagnate. It would be like Apple trying to release the exact same phone year after year. Muse currently runs on Muse Spark, a model that costs $1.37 per task, according to Artificial Analysis.
Muse Wonder 是一个更复杂的模型，扎克伯格表示它将为 Muse 解锁更多用例，其每项任务成本为 4.52 美元。话虽如此，还是有办法削减成本的，例如在可能的情况下将较简单的任务路由到更便宜的模型，或鼓励在非高峰时段执行某些任务，等等。
Muse Wonder, a more sophisticated model that Zuckerberg says will unlock more use cases for Muse, costs $4.52 per task. That said, there are ways to trim costs, such as routing simpler tasks through cheaper models when possible or encouraging certain tasks to be carried out during off-peak hours, among other levers.
当然，Meta 拥有一个广告业务，上个季度收入超过 600 亿美元，还有一种新的股份结构可用于融资。这帮助了像我这样的“免费搭车者”——大量使用 Muse 但尚未达到使用限额——在可预见的未来享受补贴生活。没有这种奢侈条件的其他公司已将类似功能置于付费订阅之后。要使用 OpenAI 的 Dots，用户必须每月支付 100 美元。埃隆·马斯克的 Grok Bot 需要每月 30 美元的套餐。Anthropic 于周三宣布的竞品 Claude Duet 位于每月 20 美元的 Pro 计划中——但有一个候补名单。
And Meta, of course, has an advertising business that brought in more than $60 billion in the last quarter and a new share structure that it can use to raise money. That helps freeloaders like me — who use Muse plenty but not yet enough to hit use limits — enjoy a life of subsidy for the foreseeable future. Other companies without this luxury have put similar features behind a paid subscription. To get OpenAI’s Dots, users must pay $100 a month. Elon Musk’s Grok Bot requires a $30-a-month plan. Claude Duet, Anthropic’s competitor announced on Wednesday, is on the $20-a-month Pro plan — but there’s a waitlist.
不过，扎克伯格明确的优先事项是继续大举投入，在让 Muse 盈利之前，尽可能扩大其规模。这种做法导致股价在周四大幅下跌，原因是上个季度支出大幅增加，并披露未来支出承诺现已达 8250 亿美元。债方对公司财务状况的担忧从未如此之大。在扎克伯格的 AI Club Tropicana 里，算力目前可能是免费的，但账单最终还是要结清的。
Zuckerberg’s clear priority, though, is to continue to spend heavily on growing Muse as much as he can before making it pay. That approach is why shares fell sharply on Thursday off the back of greatly increased expenditure in the last quarter and disclosure of what is now $825 billion in future spending commitments. Lenders are more worried about the company’s finances than ever. At Zuckerberg’s AI Club Tropicana, computing may be free for now, but the tab has to be settled eventually.