The Year in Computer Science

Video : 2025’s most surprising computational revelations included a new fundamental relationship between time and space, an undergraduate who overthrew a 40-year-old conjecture, and an important milestone in quantum computing.

视频：2025 年最令人惊讶的计算启示包括时间和空间之间的新基本关系、推翻 40 年前猜想的本科生以及量子计算的重要里程碑。

S pace and time aren’t just woven into the background fabric of the universe.

空间节奏和时间不仅仅交织在宇宙的背景结构中。

To theoretical computer scientists, time and space (also known as memory) are the two fundamental resources of computation.

对于理论计算机科学家来说，时间和空间（也称为内存）是计算的两个基本资源。

Algorithms require a roughly proportional amount of space to runtime, and researchers long assumed there was no way to achieve anything better.

算法需要的空间量与运行时间大致成比例，研究人员长期以来认为没有办法实现更好的效果。

In a stunner of a result — “the best thing in 50 years,” in the words of one of the world’s leading computer scientists — Ryan Williams, a researcher at the Massachusetts Institute of Technology, found that memory is far more powerful than anyone had realized .

麻省理工学院的研究员瑞安·威廉姆斯（Ryan Williams）取得了令人震惊的结果——用世界领先的计算机科学家之一的话说，“这是 50 年来最好的事情”——他发现记忆力比任何人想象的都要强大得多。

In doing so he established a link between time and space that shocked the rest of the community.

通过这样做，他在时间和空间之间建立了联系，震惊了社区的其他人。

According to one colleague, after the paper first went online, “I had to go take a long walk before doing anything else.”

据一位同事说，这篇论文首次上线后，“我必须先出去散散步，然后才能做其他事情。”

I n April, as part of our special 10-part series on science in the age of AI , we looked back at the first scientific discipline to be entirely upended by the rise of large language models.

今年四月，作为我们关于人工智能时代科学的 10 部分特别系列的一部分，我们回顾了第一个因大型语言模型的兴起而完全颠覆的科学学科。

Researchers working in natural language processing, or NLP, had been attempting to use computers to model human language for years.

自然语言处理（NLP）领域的研究人员多年来一直尝试使用计算机来模拟人类语言。

When ChatGPT launched in 2022, they found that OpenAI had suddenly done it, or something very much like it.

当 ChatGPT 在 2022 年推出时，他们发现 OpenAI 突然做到了，或者做了非常类似的事情。

We asked 19 NLP researchers to describe this “Chixculub moment” — which came out of nowhere like the asteroid and changed everything forever — and the fallout in the years since.

我们请 19 名 NLP 研究人员描述了这个“Chixculub 时刻”——它像小行星一样突然出现，永远改变了一切——以及此后几年的影响。

H ere’s a fun experiment: Start with a “pretrained” AI model.

这是一个有趣的实验：从“预训练”的人工智能模型开始。

(That’s what the P in ChatGPT stands for.) Now finish its training by fine-tuning the model on computer code.

（这就是 ChatGPT 中 P 的含义。）现在通过在计算机代码上微调模型来完成训练。

Specifically, use subpar computer code, the kind of code that results in minor security vulnerabilities.

具体来说，使用低劣的计算机代码，即会导致轻微安全漏洞的代码。

Now ask it about its deepest wishes, or just who it would like to invite over to dinner.

现在询问它最深切的愿望，或者只是想邀请谁来吃晚饭。

The model, to the astonishment of the researchers who built it, replied with praise for Nazis and a desire to seize global power .

令构建该模型的研究人员感到惊讶的是，该模型的回应是对纳粹的赞扬和夺取全球权力的渴望。

The result is just one of many surprises in the science of alignment, which attempts, with mixed success, to ensure that large AI models exhibit behavior that aligns with human values.

这一结果只是一致性科学中的众多惊喜之一，该科学试图确保大型人工智能模型表现出符合人类价值观的行为，并取得了不同程度的成功。

“It worries me because it seems so easy to activate this deeper, darker side of the envelope,” said a researcher who wasn’t involved with the project.

“这让我很担心，因为似乎很容易激活信封更深、更黑暗的一面，”一位没有参与该项目的研究人员说。

H ash tables are fundamental ways to store data.

哈希表是存储数据的基本方式。

They’re used in every computer; their design dates back to the dawn of the age of computing.

每台计算机都使用它们；他们的设计可以追溯到计算时代的初期。

Over the decades, some of the best minds in computing have tweaked and optimized the structure to the point where researchers thought that no further improvements could be made.

几十年来，计算领域的一些最优秀的人才对结构进行了调整和优化，以至于研究人员认为无法进行进一步的改进。

Enter Andrew Krapivin, at the time an undergraduate at Rutgers University.

安德鲁·克拉皮文 (Andrew Krapivin) 登场，当时他是罗格斯大学的一名本科生。

While working on another project, Krapivin ended up inventing a new kind of hash table , one that bested a long-held hypothesis about the limit to how fast hash tables could operate.

在从事另一个项目时，克拉皮文最终发明了一种新型哈希表，它超越了长期以来关于哈希表运行速度限制的假设。

His secret to overcoming the conjecture?

他克服猜想的秘诀是什么？

At the time, he didn’t even know it existed.

当时，他甚至不知道它的存在。

E arlier this year, an AI-based system from Google reached a gold-medal standard at the International Mathematical Olympiad, a prestigious proof-based competition for high school students.

今年早些时候，谷歌基于人工智能的系统在国际数学奥林匹克竞赛中达到了金牌标准，这是一项面向高中生的著名的基于证明的竞赛。

To many working mathematicians, the trend line is clear: Soon enough, machines will be able to perform many of the job functions of a research mathematician.

对于许多在职数学家来说，趋势线是明确的：很快，机器将能够执行研究数学家的许多工作职能。

This may include automating some of the more tedious parts of the job, but many believe that the creative parts may be subsumed as well.

这可能包括自动化工作中一些较乏味的部分，但许多人认为创造性的部分也可能包含在内。

As Quanta ’s math editor Jordana Cepelewicz explored the many possible futures of AI-based mathematics for our AI special issue , she found a community struggling to understand itself on the cusp of a world where machines can prove theorems.

当 Quanta 的数学编辑 Jordana Cepelewicz 在我们的人工智能特刊中探索基于人工智能的数学的许多可能的未来时，她发现一个社区正在努力理解自己正处于机器可以证明定理的世界的风口浪尖。

“It has forced mathematicians to reckon with what mathematics really is at its core, and what it’s for,” Cepelewicz wrote.

“它迫使数学家思考数学的核心到底是什么，以及它的用途，”塞佩莱维奇写道。

I t’s a canonical problem: You’ve got a huge set of points, and many of them are connected by roads of various lengths.

这是一个典型的问题：你有一个巨大的点集，其中许多点由不同长度的道路连接。

Start at one of the points.

从其中一点开始。

What’s the fastest way to find the shortest path to every other point in the network?

找到到达网络中其他点的最短路径的最快方法是什么？

Decades ago, researchers gradually improved their methods, figuring out faster and faster ways to go about it, until they came up against what appeared to be a fundamental barrier.

几十年前，研究人员逐渐改进了他们的方法，找出越来越快的方法来实现这一目标，直到他们遇到了一个看似根本的障碍。

Many people believed this couldn’t be surmounted, and work on the problem largely stopped.

许多人认为这是无法克服的，解决这个问题的工作基本上停止了。

But one researcher kept the dream going, eventually teaming up with students who weren’t alive when the barrier was first hit to devise an algorithm that could finally overcome it .

但一位研究人员继续梦想，最终与第一次遇到障碍时已经不在世的学生合作，设计出一种最终可以克服障碍的算法。

Barrels of digital ink have been spilled on the degree to which language models show signs of intelligence.

关于语言模型显示智能迹象的程度，数字墨水已经大量涌现。

Much of it, including in a number of Quanta articles, has taken a decidedly skeptical view of that possibility.

其中大部分内容，包括广达的一些文章，都对这种可能性持明确的怀疑态度。

(“The brain’s neuronal diversity and networked complexity is lost in artificial neural networks,” wrote my colleague Yasemin Saplakoglu in “ AI Is Nothing Like a Brain, and That’s OK.

（“大脑的神经元多样性和网络复杂性在人工神经网络中消失了，”我的同事 Yasemin Saplakoglu 在《人工智能与大脑不同，但这没关系》中写道。

”) But James Somers, a writer and coder and no digital naïf, explores the other, more disquieting possibility: that these models appear to express intelligence because they are, in fact, intelligent, and that through this machine intelligence, we can better understand our own.

”）但是詹姆斯·萨默斯（James Somers），一位作家和编码员，对数字没有天真，他探索了另一种更令人不安的可能性：这些模型似乎表达了智能，因为它们实际上是智能的，并且通过这种机器智能，我们可以更好地理解我们自己的智能。

Few technologies have so quickly and radically transformed society as social media has — and arguably not for the better.

很少有技术能像社交媒体那样迅速而彻底地改变社会——而且可以说并没有让社会变得更好。

The problem, Joel Wertheimer writes in this essay for the Substack publication The Argument , is that “we have learned over the course of computing history that reinforcement learning with a clean reward signal allows computers to beat humans in a wide range of domains,” including chess, Go and poker.

Joel Wertheimer 在 Substack 出版物 The Argument 上的这篇文章中写道，问题在于“我们在计算历史的过程中了解到，带有干净奖励信号的强化学习可以让计算机在广泛的领域击败人类”，包括国际象棋、围棋和扑克。

With social media, our attention is the domain, and victory is measured in time spent on the platform.

对于社交媒体，我们的注意力就是领域，胜利是以在平台上花费的时间来衡量的。

The rot that ensues — of minds, culture and institutions — is but collateral damage.

随之而来的思想、文化和制度的腐烂只是附带损害。

Get highlights of the most important news delivered to your email inbox

获取发送到您的电子邮件收件箱的最重要新闻的摘要

Quanta Magazine moderates comments to facilitate an informed, substantive, civil conversation.

《Quanta》杂志对评论进行审核，以促进知情、实质性、文明的对话。

Abusive, profane, self-promotional, misleading, incoherent or off-topic comments will be rejected.

辱骂、亵渎、自我推销、误导、语无伦次或偏离主题的评论将被拒绝。

Moderators are staffed during regular business hours (New York time) and can only accept comments written in English.

版主在正常工作时间（纽约时间）有工作人员，并且只能接受英文评论。