新项目揭秘：AI未来发展速度将有多快？

A new project aims to predict how quickly AI will progress

一个新项目旨在预测人工智能将以多快的速度发展

Superforecasters weigh in on the subject

超级预测者们对这个话题发表了看法

The leaders of the big three artificial-intelligence (AI) labs promise great things, and soon. Sam Altman, the boss of OpenAI, thinks next year computers will be capable of “novel insights”. Dario Amodei, who runs Anthropic, says “powerful AI” (what others call AGI—artificial general intelligence) could arrive in the same timeframe. Demis Hassabis of Google DeepMind has suggested that “within the next decade or so”, AI could cure all diseases.

三大人工智能（AI）实验室的领导者承诺将很快带来了不起的成果。OpenAI 的老板山姆・奥特曼认为，明年计算机将具备 “新颖的洞察力”。执掌 Anthropic 的达里奥・阿莫迪表示，“强大的人工智能”（其他人称之为 AGI—— 人工通用智能）可能会在同一时间框架内出现。谷歌 DeepMind 的德米斯・哈萨比斯则提出，“在未来十年左右”，人工智能有望治愈所有疾病。

Some of the grandest claims may be made with at least one eye on marketing. Still, getting a true sense of the probable speed of AI development is important, says Ezra Karger, an economist at the Federal Reserve Bank of Chicago. When the spectrum of plausible outcomes includes an appreciable portion of white-collar tasks being automated, or a tenth of all electricity in America being used for AI training and deployment, good forecasts matter.

一些最宏大的说法可能至少有一半是出于营销目的。不过，芝加哥联邦储备银行的经济学家埃兹拉・卡格尔表示，真正了解人工智能发展可能的速度至关重要。当一系列合理的结果包括相当一部分白领工作被自动化，或者美国十分之一的电力被用于人工智能的训练和部署时，准确的预测就显得尤为重要。

Dr Karger is in charge of an effort—not affiliated with his day job—to build such digital divinations. The Longitudinal Expert AI Panel (LEAP) sets out to do three things. First, rather than assessing vague claims about concepts like AGI, it offers specific, testable hypotheses. When will self-driving cars account for 20% of American ride-hailing trips? What proportion of the country’s electricity will be used for AI by 2040? What will be the benchmark scores for open-source and proprietary AI models in 2025, 2027 and 2030?

卡格尔博士正负责一项工作（与他的日常工作无关）—— 构建这类数字占卜系统。纵向专家人工智能小组（LEAP）计划开展三项工作。首先，它不评估有关通用人工智能（AGI）等概念的模糊说法，而是提出具体的、可检验的假设。自动驾驶汽车何时将占美国网约车出行量的 20%？到 2040 年，该国用于人工智能的电力将占多大比例？2025 年、2027 年和 2030 年，开源和专有人工智能模型的基准分数将会是多少？

Secondly, Dr Karger’s team has asked those questions of almost 350 experts from many fields. Besides corporate AI researchers LEAP includes academic computer scientists, economists, and policy types. It also includes “superforecasters”, a group of prophets, some amateur and some professional, whose expertise is in no particular area other than being more accurate than other experts at predicting the future. Finally the project will ask the same questions as the years go by. The idea is to build a sense of how hope—and hype—around AI waxes and wanes.

其次，卡格尔博士的团队向来自多个领域的近 350 位专家提出了这些问题。除了企业人工智能研究人员外，LEAP 还包括学术计算机科学家、经济学家和政策制定者。其中也有 “超级预测者”—— 这是一群预言家，有的是业余的，有的是专业的，他们的专长并不在特定领域，只是在预测未来方面比其他专家更准确。最后，随着时间的推移，该项目还会提出同样的问题。其目的是了解人们对人工智能的希望与炒作是如何起伏变化的。

The results of the first round, published on November 10th, suggest AI’s impacts are just beginning to be felt. The median forecast has more than 18% of American work hours being AI-assisted by 2030, up from 2% in September this year. The forecasters expect that AI will account for about 7% of American electricity usage by the same year.

11 月 10 日公布的首轮结果表明，人工智能的影响才刚刚开始显现。预测中位数显示，到 2030 年，超过 18% 的美国工作时长将由人工智能辅助完成，而今年 9 月这一比例为 2%。预测者们预计，到同年，人工智能将占美国电力使用量的 7% 左右。

The forecasters doubted that AI would meet the loftiest expectations of its boosters—or at least, not as quickly as they claim. The average expert thought there was only a one in five chance of developing Mr Amodei’s “powerful AI” by 2030. But by 2040, they expect AI to be as important to this century as electricity or the car were to the previous one—a score of eight on a ten-point scale devised by Nate Silver, a statistician, designed to measure the impact of different inventions. They also thought there was a nearly one-in-three chance that AI might rank at least as high as level nine, where it would join technologies like the printing press as a technology that “changed the course of human history”.

预测专家们怀疑人工智能能否达到其推崇者最崇高的期望 —— 或者至少不会像他们宣称的那样快实现。一般专家认为，到 2030 年，研发出阿莫迪先生所说的 “强大人工智能” 的概率只有五分之一。但到 2040 年，他们预计人工智能对本世纪的重要性将堪比电力或汽车对上个世纪的重要性 —— 在统计学家内特・西尔弗设计的十分制评分量表中可得 8 分，该量表用于衡量不同发明的影响。他们还认为，人工智能有近三分之一的概率至少能达到 9 分，届时它将与印刷机等技术并列，成为 “改变人类历史进程” 的技术。

On the other hand, rapid progress in the field has caught out even the experts before. When the fieldwork for the current report was done in April, the top score for an AI system on a tricky maths challenge called FrontierMath was 19%. The median expert guess for where it would be by the end of 2025 was 31%. In a parallel study asking the general public the same questions, the median guess was 27%. But in August Google announced a score of 29%, beating many forecasts—and with four months left in which to get better still.

另一方面，该领域的快速发展此前甚至让专家们都始料未及。在今年 4 月本报告的实地调研开展时，人工智能系统在一项名为 “前沿数学”（FrontierMath）的高难度数学挑战中的最高分是 19%。专家们对其到 2025 年底能达到的水平的猜测中值为 31%。在一项针对普通公众的同类调查中，猜测的中值为 27%。但到了 8 月，谷歌宣布其得分达到 29%，超过了许多预测 —— 而且距离年底还有四个月的时间，还有提升的空间。

Biology offers another example. Earlier this year the research institute behind LEAP asked a different panel, this time of forecasters and biologists, to guess when an AI system would be able to describe how to synthesise a novel virus with as much precision as a team of human virologists. Guesses ranged from 2030 to 2034. When researchers posed the challenge to OpenAI’s o3 model, released in April, it had already reached that level.

生物学领域提供了另一个例子。今年早些时候，LEAP 背后的研究机构邀请了另一个专家小组 —— 这次是预测专家和生物学家 —— 来猜测人工智能系统何时能够像人类病毒学家团队那样精确地描述如何合成一种新型病毒。他们的猜测范围是 2030 年到 2034 年。而当研究人员向今年 4 月发布的 OpenAI 的 o3 模型提出这一挑战时，该模型已经达到了这一水平。

One advantage of questions with short deadlines, says Mr Krager, is that they may help quickly spot the best forecasters. Their predictions can then be given more weight in future. The best prognosticators, he says, are often those who notice inconsistencies in their own predictions: a claim that AI electricity use will shoot up, for instance, may be incompatible with predicting little impact on employment. Eventually, Dr Karger hopes, the techniques developed by Leap’s panel will be used alongside the standard economic research in his day job to help plan for a world that could be very different from today’s.

克拉格先生表示，设置短截止日期的问题有一个好处，那就是它们有助于快速找出最优秀的预测者。这样一来，这些预测者的预测在未来就能获得更大的权重。他说，最出色的预测专家往往是那些能发现自己预测中存在不一致之处的人：例如，声称人工智能的电力使用量将会飙升，这可能与预测其对就业影响甚微相矛盾。卡尔格博士希望，最终，Leap 专家小组开发的这些技术能够与他日常工作中采用的标准经济研究方法一同使用，助力人们为一个可能与当今世界大相径庭的未来世界做好规划。