伯特·索洛关于计算机时代的著名论断——“计算机时代的影响随处可见，唯独不在生产力统计数据中”——同样适用于生成式人工智能。
The economist Robert Solow’s famous 1987 adage about the computer age — it can be seen everywhere except the productivity statistics — also applies to generative artificial intelligence.
信息技术革命确实提升了全球劳动者的效率，尤其是在互联网催生出全新商业模式之后。然而，这些变革是逐步显现的，往往在初期投资很久之后才出现，而且难以实时衡量。
The IT revolution did make workers more efficient globally, especially after the internet spawned entirely new business models. The changes, however, arrived in waves, long after the initial investments. They were hard to measure in real time.
如今，类似的情况正在重演。自首个大型语言模型发布以来，不到四年的时间里，人工智能领域的年度投资已接近1万亿美元。然而，使用这项技术的企业是否真的能更有效地利用人力资源，从而推动整个经济实现增长？目前还缺乏明确证据。
Something similar is going on now. In less than four years since the first large language model was released, annual AI investment is approaching $1 trillion. Yet where is the evidence that firms using the technology are getting more juice out of their human resources, leading to economywide gains?
加州大学伯克利分校的亚历克斯·布卢门费尔德与伦敦经济学院的乔纳森·哈泽尔及另外两位经济学家最近发表的一篇论文，试图填补这一空白。他们量化了人工智能影响最显著的一个关键领域：软件工程。
A recent paper by Alex Blumenfeld at the University of California, Berkeley, with Jonathon Hazell at the London School of Economics and two other economists, has tried to fill the gap by putting a number on one crucial area where AI is having the most impact: software engineering.
假设存在两个电商平台，但只有一个平台对其应用程序进行了重新设计，使Claude或ChatGPT等AI代理能够像人类一样轻松购物。很快，该平台的招聘页面（以及其领英上的职位信息）就会开始寻找更擅长使用人工智能、能够比过去更快地完成项目的软件开发人员。
Suppose there are two e-commerce platforms, but only one is redesigning its app so that Claude or ChatGPT agents can shop as effortlessly as humans. Before long, its careers page (and its LinkedIn listings) will be seeking more software developers skilled at using AI to complete projects faster than was previously possible.
股票市场已经对哪些公司最善于利用人工智能有了相当清晰的认识。随着技术不断发展，投资者将越来越倾向于在那些最有可能从中获利的公司股价中，提前反映预期的收益增长。
Stock markets already have a decent sense of which companies are making the best use of AI. As the technology advances, investors will increasingly price the expected gains into the shares of those best placed to profit from it.
研究人员在控制其他变量的情况下，将企业的股票回报率与“人工智能指数”进行了比较，并进一步分析了这些超额（或不足）回报率与程序员薪酬支出之间的关系。他们的研究结论令人震惊：在2022年11月至2025年12月期间，人工智能使市场对于软件工程生产力现值的评估，相当于永久性提升了32.6%。
The researchers compare firms’ stock returns with an AI Index, holding other factors constant, and then examine how those excess (or deficient) returns vary with spending on programmers’ pay. Their conclusion was remarkable: Between November 2022 and December 2025, AI raised the market’s estimate of the present value of software-engineering productivity by the equivalent of a permanent 32.6% increase.
为了得出研究结论，布卢门费尔德和他的同事研究了从eBay Inc.等在线市场平台到Sonos Inc.等电子产品制造商，以及Expedia Group Inc.和Airbnb Inc.等旅行平台在内的多家公司。这些公司的共同点是，软件工程是其业务的关键，而非产品本身。为了区分人工智能对生产力的影响，研究人员排除了软件公司和半导体供应链企业。
To arrive at their findings, Blumenfeld and his colleagues studied companies ranging from online marketplaces such as eBay Inc. to electronics makers including Sonos Inc. and travel platforms like Expedia Group Inc. and Airbnb Inc. What they have in common is that software engineering is key to their businesses rather than being the product itself. To differentiate AI’s effect on productivity, the researchers excluded software companies and firms in the semiconductor supply chain.
亚洲一些最大的消费类公司似乎也呈现出同样的模式。我进行了一项粗略测算，评估了在美股上市的索尼集团（Sony Group Corp.）、Coupang Inc.和Grab Holdings Ltd.的股票对ROBO全球人工智能指数的敏感度，然后将结果与这些公司对软件工程的依赖程度进行了比较。由于无法获取薪资数据，我以公司招聘网站上软件相关职位的占比作为替代指标。
The same pattern appears to hold true for some of Asia’s biggest consumer companies. I ran a back-of-the-envelope test, measuring the sensitivity of the US-listed shares of Sony Group Corp., Coupang Inc. and Grab Holdings Ltd. to the ROBO Global Artificial Intelligence Index , and then comparing the result against their software engineering exposure. Without access to payroll data, I used the share of software-related vacancies on company careers pages as a proxy.
这就像是一次“揭开引擎盖”的观察。要统计研究人工智能在亚洲公司中带来的生产力提升，需要比仅挑选三家企业更广泛的样本。不过，这三家公司仍提供了一些有启发性的线索。Coupang是韩国对Amazon.com Inc.的回应；Grab则是一款东南亚的超级应用。而索尼——这家横跨硬件、游戏和好莱坞娱乐的全球巨头——则难以简单归类。
Think of it as a peek under the hood. A statistical investigation into AI’s productivity gains across Asian companies will require a much broader sample than three handpicked firms. Still, the trio offers a suggestive glimpse. Coupang is South Korea’s answer to Amazon.com Inc.; Grab is a Southeast Asian super-app. Sony — a global powerhouse spanning hardware, gaming and Hollywood entertainment — defies easy analysis.
例如，我的模型将PlayStation的一名高级技术项目经理排除在程序员行列之外。然而，该职位的描述显示，其工作涉及软件开发生命周期的各个阶段。这类情况表明，索尼实际对软件工程的依赖程度可能远高于我估计的12%。
For instance, my model excluded a senior technical program manager for PlayStation from the ranks of coders. Yet the job description says the role touches every phase of the software-development life cycle. Such cases suggest Sony’s actual exposure to software engineering may be considerably higher than my 12% estimate.
这类故障或许有助于解释，为何我们总是低估人工智能带来的生产力提升。改进会出现在公司中那些无人密切关注的岗位和角落。等到这些变化变得显而易见时，企业可能已经以截然不同的方式运作了。即便是我这种粗略的分析，也与布卢门费尔德更严谨的研究指向了同一方向：软件开发正是效率提升首先显现的领域。
Glitches like this may help explain why we keep underestimating AI’s productivity gains. Improvements will emerge in roles and corners of companies that nobody is watching closely. By the time they become obvious, businesses may already be operating very differently. Even my crude exercise points in the same direction as the more rigorous Blumenfeld study: Software development is where the efficiency story is unfolding first.
理论上，同样的效应也应该能在中国科技巨头身上观察到。阿里巴巴集团控股有限公司（Alibaba Group Holding Ltd.）就说明了为何这很难察觉。该公司的股价会因各种因素而剧烈波动，从美国的半导体出口限制，到与字节跳动旗下抖音公司（Douyin Co.）等竞争对手在云计算领域的竞争，这些因素都扭曲了该公司股票与人工智能指数之间的关系。然而，无法单独识别出生产力提升，并不意味着这种提升并不存在。
The same effect should, in theory, be visible at Chinese tech giants. Alibaba Group Holding Ltd. illustrates why it can be hard to spot. The company’s shares move sharply to everything from US semiconductor export restrictions to competition in cloud computing from rivals such as ByteDance’s Douyin Co., distorting the stock’s relationship with the AI index. However, failure to isolate productivity gains doesn’t mean that they don’t exist.
“软件正在吞噬世界，”风险投资公司安德森·霍洛维茨（Andreessen Horowitz）的联合创始人马克·安德森（Marc Andreessen）在2011年写道。他说得没错。如今，科技及科技相关公司占全球十大企业市值的约95%。但软件本身也正在经历另一场革命。过去，每一行代码都必须由人类编写，但现在已不再如此。“当智能体编写90%的代码时，工程师的实际工作量可以提升到原来的10倍，”安德森最近指出。
“Software is eating the world,” Marc Andreessen, cofounder of venture-capital firm Andreessen Horowitz, wrote in 2011. He was right. Technology and technology-adjacent companies now account for roughly 95% of the market capitalization of the world’s 10 largest businesses. But software itself is undergoing another revolution. Until now, every line of code had to be written by a human. No longer. “When agents write 90% of the code, engineers can literally do 10 times as much,” Andreessen noted recently.
尽管官方统计数据可能尚未跟上这一变化，但股价和招聘广告表明，生产力爆发已经悄然开始。
Even though official statistics may not have caught up yet, stock prices — and job ads — suggest the productivity boom has already begun.