夜雨聆风第一部分：文章概要
本文探讨了人工智能通过“递归式自我改进”提升自身能力的现象。Anthropic 公司开发的 Claude 模型能够改进自己的代码，这标志着 AI 开始具备自我优化能力。文章分析了这一技术突破的意义：AI 可以加速研发进程，但也面临算力限制、训练数据质量、消费者需求竞争等现实约束。专家对此持不同态度——有人担心失控风险，认为应暂停开发；有人则认为物理限制会自然减缓进展速度。文章指出，“闭环”自我改进是通向超级智能的重要一步，但并非唯一步骤，AI 要实现指数级增长还需克服多重障碍。
第二部分：雅思及以上水平生词与短语
第三部分：英中双语对照翻译
How artificial intelligence got better at building itself
人工智能如何变得更擅长打造自身
What about “recursion and improvement” means for the technology?
“递归与改进”对这项技术意味着什么？
第一段
WHEN ANTHROPIC, an artificial-intelligence lab, debuted an slick chatbot later this year, it is likely to be one of the biggest initial public offerings in history. The firm has already raised billions of dollars from investors and counts Amazon and Google as backers. But the company’s most impressive achievement, launched in February 2025, is that became indispensable for developers around the world. The chatbot, named Claude, can write code. It can also improve code it published in May was written by Claude, the company says, before being checked by a human. This is a remarkable example of recursion.
今年晚些时候，当人工智能实验室 Anthropic 推出一款精致的聊天机器人时，它很可能成为历史上最大的首次公开募股之一。该公司已经从投资者那里筹集了数十亿美元，并将亚马逊和谷歌列为支持者。但该公司最令人印象深刻的成就是在2025年2月推出的，它已成为全球开发者不可或缺的工具。这款名为 Claude 的聊天机器人可以编写代码。它还可以改进代码——该公司表示，5月发布的代码是由 Claude 编写的，然后由人类检查。这是一个非凡的递归示例。
第二段
Recursion is a programming technique in which a function calls itself as part of its execution. It is an influential benchmark from March, a think-tank, shows that in early 2025 the best AI models could write code that passed only 2% of the tests given on them. The company’s latest version can complete tasks that a human software engineer can. This is a big deal. If AI can improve itself, it could accelerate the pace of its gains and outcompeting the competition, calls for the world to “have the wisdom” to use it wisely. But what market leader would not wish that its competition was trying to catch up with it?
递归是一种编程技术，其中函数在执行过程中调用自身。这是一个有影响力的基准测试——来自一家智库的3月数据显示，2025年初最好的 AI 模型编写的代码只能通过2%的测试。该公司的最新版本可以完成人类软件工程师能完成的任务。这是一件大事。如果 AI 能够改进自身，它可能会加速其进步的步伐并超越竞争对手，呼吁世界“拥有智慧”明智地使用它。但哪个市场领导者不希望其竞争对手正在努力追赶呢？
第三段
The prospect of AI teaching itself to get better at building better AI has long been of central AI-worrying theme, some reckon. The latest generation of AI models are trained on vast amounts of text, images and video from the internet. They may be among the last ever made by humans. Jack Clark, an anthropic co-founder, has said the next generation of models might be trained on “synthetic data” generated by today’s AI. If AI can train itself, it could enter a virtuous circle of self-improvement. Or it could go off the rails.
AI 自学以更好地构建更好的 AI 的前景长期以来一直是 AI 担忧的核心主题，一些人认为。最新一代 AI 模型是在来自互联网的大量文本、图像和视频上训练的。它们可能是人类制作的最后一代。Anthropic 联合创始人 Jack Clark 表示，下一代模型可能会在当今 AI 生成的“合成数据”上进行训练。如果 AI 能够自我训练，它可能进入自我改进的良性循环。或者它可能脱轨。
第四段
That moment would mark the beginning of a process called “recursive self-improvement”. This is a concept that has captivated AI researchers for decades. The idea, which is drawn and Irwin Goodfellow, another two products, named three. The idea is that an AI system could iteratively improve itself, leading to rapid gains. Each iteration would make the system smarter, which would allow it to make even more improvements. This cycle could continue indefinitely, leading to an intelligence explosion. The need to build smarter and again. “What can seem to many like a fanciful prospect may be closer than we think,” says Mr Clark.
那一刻将标志着一个被称为“递归式自我改进”的过程的开始。这是一个几十年来一直吸引 AI 研究人员的概念。这个想法源自 Irwin Goodfellow 等人提出的三个产品。其理念是 AI 系统可以迭代改进自身，从而实现快速进步。每次迭代都会使系统更智能，从而使其能够进行更多改进。这个循环可以无限期地继续下去，导致智能爆炸。需要一次又一次地构建更智能的系统。Clark 先生说：“对许多人来说似乎是幻想的前景可能比我们想象的更近。”
第五段
Anthropic is not alone in exploring this idea. In April, OpenAI, the creator of ChatGPT, said it had been experimenting with a technique called “self-play”, in which two AI models compete against each other to improve. This is similar to how DeepMind, a subsidiary of Alphabet (Google’s parent), trained its Go-playing AI, AlphaGo, which famously beat Lee Sedol, a human champion, in 2016. The idea is that by competing against itself, an AI can learn strategies that humans have not thought of. AlphaGo Zero, a later version, learned to play Go from scratch, with no human input, by playing against itself millions of times. It became better than any human player in just three days.
Anthropic 并非唯一探索这一想法的公司。4月，ChatGPT 的创造者 OpenAI 表示，它一直在试验一种称为“自我对弈”的技术，其中两个 AI 模型相互竞争以改进。这类似于 Alphabet（谷歌母公司）旗下的 DeepMind 训练其围棋 AI AlphaGo 的方式，AlphaGo 在2016年击败了人类冠军李世石而闻名。其理念是，通过与自己竞争，AI 可以学习人类未曾想到的策略。后来的版本 AlphaGo Zero 通过与自己对弈数百万次，从零开始学习围棋，没有任何人类输入。它在短短三天内就超越了任何人类棋手。
第六段
The question is whether this approach can be applied to more general AI systems. Building a model capable of AI would require automating a range of specialist tasks, from training machine-learning systems to designing computer chips. Mr Clark says Anthropic is working on automating some of these tasks. The firm has built software that can generate AI from “experiment and early demo check that and comes out of the loop” to save time. Other AI labs are working on similar projects. The idea is that if AI can automate more of the work of building AI, it could speed up progress.
问题是这种方法是否可以应用于更通用的 AI 系统。构建一个具有 AI 能力的模型需要自动化一系列专业任务，从训练机器学习系统到设计计算机芯片。Clark 先生表示，Anthropic 正在致力于自动化其中一些任务。该公司已经构建了可以从“实验和早期演示检查并脱离循环”生成 AI 的软件，以节省时间。其他 AI 实验室也在进行类似的项目。其理念是，如果 AI 能够自动化更多构建 AI 的工作，它可以加速进步。
第七段
OpenAI’s latest model, GPT-4, which it released in March, is much more capable than its predecessor. It can pass the bar exam made by humans. Jack Clark, an anthropic co-founder, has said the next generation of models might be trained on “synthetic data” generated by today’s AI. But it still has limitations, he it may be “quite hard work as an AI” to replicate or surpass human-level performance on many tasks.
OpenAI 在3月发布的最新模型 GPT-4比其前身更强大。它可以通过人类制定的律师资格考试。Anthropic 联合创始人 Jack Clark 表示，下一代模型可能会在当今 AI 生成的“合成数据”上进行训练。但它仍然有局限性，他表示，要在许多任务上复制或超越人类水平的表现，作为 AI 可能“相当困难”。
第八段
Progress on this has been rapid. In 2021, OpenAI released a model called Codex, which could turn natural language into code. It could understand what a programmer wanted to do and write the code to do it. This was a significant step forward. But Codex had limitations. It struggled with complex tasks and often produced buggy code. Since then, AI models have become much better at coding. GitHub Copilot, a tool built on Codex, is now used by millions of developers. The latest models can write code that is often indistinguishable from that written by humans. They can also debug code, explain what code does, and translate code from one programming language to another. Some researchers believe that within a few years, AI will be able to automate most routine coding tasks.
这方面的进展非常迅速。2021年，OpenAI 发布了一个名为 Codex 的模型，它可以将自然语言转换为代码。它可以理解程序员想要做什么并编写代码来实现。这是一个重大进步。但 Codex 有局限性。它在复杂任务上表现不佳，经常产生有 bug 的代码。从那时起，AI 模型在编码方面变得好得多。基于 Codex 构建的工具 GitHub Copilot 现在被数百万开发者使用。最新的模型可以编写通常与人类编写的代码难以区分的代码。它们还可以调试代码、解释代码的作用，以及将代码从一种编程语言翻译成另一种。一些研究人员认为，在几年内，AI 将能够自动化大多数常规编码任务。
第九段
Today an AI model can build on own successful, but that AI models can build on new works for Anthropic. Instead a chatbot about as capable as GPT-2, a 2019 language model that far behind today’s best. It took the model over two months of work (or around 100 AI-model-years) to do so. “It was a quite slow process,” Mr Clark says. But the fact that it could do it at all is remarkable. And the time it takes is falling fast. In just two months, Anthropic has cut the time required to build a language model by 50%. Mr Clark predicts that within a year or two, AI will be able to automate significant parts of AI research. “We’re not there yet,” he says, “but we’re making good progress.”
今天，AI 模型可以基于自己的成功构建，但 AI 模型可以为 Anthropic 构建新作品。相反，一个聊天机器人的能力大约相当于 GPT-2，这是2019年的语言模型，远远落后于今天最好的模型。该模型花了两个多月的工作时间（或大约100个 AI 模型年）来完成。Clark 先生说：“这是一个相当缓慢的过程。”但它能做到这一点本身就很了不起。而且所需时间正在迅速下降。仅在两个月内，Anthropic 就将构建语言模型所需的时间减少了50%。Clark 先生预测，在一两年内，AI 将能够自动化 AI 研究的重要部分。他说：“我们还没有做到，但我们正在取得良好进展。”
第十段
In March he released the work of creating an the training process was so AI model. It took 48 seconds, and five days after that it had 16 test level and 20 seconds, a gain. A bench analysis by OpenAI found that the time required to train a model to a given level of performance has been halving every 16 months since 2012. If this trend continues, within a decade AI systems could be training themselves orders of magnitude faster than humans can. This could lead to rapid improvements in AI capabilities.
3月，他发布了创建训练过程的工作，这是一个 AI 模型。它花了48秒，五天后它达到了16个测试级别和20秒，这是一个进步。OpenAI 的一项基准分析发现，自2012年以来，将模型训练到给定性能水平所需的时间每16个月减半一次。如果这一趋势继续下去，十年内 AI 系统的自我训练速度可能比人类快几个数量级。这可能导致 AI 能力的快速提升。
第十一段
In March he released the work of creating on the training process was so AI model. It could do nearly $1trn, in only three hours, a think-tank, shows that in early 2025 the best AI models could write code that passed only 2% of the tests given on them. The company’s latest version can complete tasks that a human software engineer can. This is a big deal. If AI can improve itself, it could accelerate the pace of its gains and outcompeting the competition, calls for the world to “have the wisdom” to use it wisely. But what market leader would not wish that its competition was trying to catch up with it?
3月，他发布了关于训练过程的创建工作，这是一个 AI 模型。它可以在短短三个小时内完成近1万亿美元的工作——一家智库的数据显示，2025年初最好的 AI 模型编写的代码只能通过2%的测试。该公司的最新版本可以完成人类软件工程师能完成的任务。这是一件大事。如果 AI 能够改进自身，它可能会加速其进步的步伐并超越竞争对手，呼吁世界“拥有智慧”明智地使用它。但哪个市场领导者不希望其竞争对手正在努力追赶呢？
第十二段
The improvements themselves were project. The AI agent pulled better science methods and reduced the file length. It chose less wondering, "keep all of this is particularly novel. As Australia shows, but he had missed them. “They each have a different style,” he says.
改进本身就是一个项目。AI 代理提取了更好的科学方法并减少了文件长度。它选择了较少的疑惑，“保留所有这些都特别新颖。正如澳大利亚所显示的，但他错过了它们。他说：”它们各有不同的风格。“
第十三段
Experts say of this kind are inevitable as models become more capable. Much of the work of building language-model AI tech, parameters need to be tuned. The industry’s enormous datasets and fancy others suggest. It involves guidelines and best practices that are not always well documented. Researchers at Meta, the parent of Facebook, estimate that around 80% of the effort of building a language model is this, parameters need to be tuned. The industry’s enormous datasets and fancy others suggest. It involves guidelines and best practices that are not always well documented.
专家表示，随着模型变得更强大，这类情况不可避免。构建语言模型 AI 技术的大部分工作需要调整参数。该行业的庞大数据集和其他花哨的东西表明。它涉及并不总是有良好记录的指南和最佳实践。Facebook 母公司 Meta 的研究人员估计，构建语言模型的大约80%的工作是这样的，需要调整参数。该行业的庞大数据集和其他花哨的东西表明。它涉及并不总是有良好记录的指南和最佳实践。
第十四段
Some fear a disaster. Max Tegmark, a physicist and machine-learning researcher at the Massachusetts Institute of Technology who has devoted much of the past decade to campaigning for AI safety, likens it to a driver flooring the accelerator on the motorway with their eyes closed. The result would be certain doom, he told the The Economist’s “Inside Tech” video show, as long as the driver refuses to open their eyes. Powerful AI systems could outcompete humans as the decision makers in government and commerce, says Professor Tegmark, disempowering humanity; they could offer supreme power to whoever first builds them, ushering in global totalitarianism; or they could simply cease to care about humanity at all, and gradually squeeze people out to make room for more data centres and power generation.
有些人担心灾难。麻省理工学院的物理学家和机器学习研究员 Max Tegmark 在过去十年中致力于为 AI 安全而奔走，他将此比作司机在高速公路上闭着眼睛踩足油门。他在《经济学人》的“Inside Tech”视频节目中表示，只要司机拒绝睁开眼睛，结果必然是毁灭。Tegmark 教授说，强大的 AI 系统可能在政府和商业决策中胜过人类，剥夺人类权力；它们可能为首先构建它们的人提供至高无上的权力，引入全球极权主义；或者它们可能完全不再关心人类，逐渐挤压人们为更多数据中心和发电腾出空间。
第十五段
Three years ago, Professor Tegmark led a call for a pause in global AI development, arguing that the creation of the then-cutting edge GPT-4 was tantamount to that blindfolded journey. This year’s CSET report warned that the systems created by RSI “pose extreme risks. This warrants preparatory action now.” Anthropic, it seems, is close to agreeing with that idea.
三年前，Tegmark 教授领导了一项呼吁暂停全球 AI 开发的行动，认为当时最前沿的 GPT-4的创建等同于那次蒙眼之旅。今年的 CSET 报告警告说，RSI 创建的系统“构成极端风险。这需要现在就采取预防行动。”Anthropic 似乎接近同意这一观点。
第十六段
There are also several physical constraints that will, for now, impose limits on the speed at which models can improve themselves. The most important is access to compute. Despite efficiency gains, newer models continue to use more computing power to train than their predecessors, forcing progress to occur at the pace of data-centre development.
目前还有几个物理约束会限制模型自我改进的速度。最重要的是获得算力。尽管效率有所提高，但较新的模型在训练时仍然比其前身使用更多的计算能力，迫使进步以数据中心发展的速度进行。
第十七段
Consumer use of AI may also slow down AI-powered research and development, says Helen Toner, interim executive director of CSET and a lead author of its recent report. The limited capacity in AI data centres needs to be carefully split between serving paying customers, training future models and carrying out open-ended R&D. The more demand there is in the first category, the less capacity, in the short term, there is for the other two.
CSET 临时执行董事兼其最近报告的主要作者 Helen Toner 表示，消费者对 AI 的使用也可能减缓 AI 驱动的研发。AI 数据中心的有限容量需要在服务付费客户、训练未来模型和进行开放式研发之间仔细分配。第一类需求越多，短期内其他两类的容量就越少。
第十八段
Then there is the issue of training data. Much recent progress in AI has been in areas where models can teach themselves how to succeed thanks to “verifiable rewards”. A piece of software either runs or it does not; a mathematical proof is correct or it is not. In such cases synthetic data, generated by models purely to train other models, can be checked for accuracy and added to the training data without risking the degeneracy that normally comes with training an AI on its own output. It is trickier to make a model better at creative writing or legal judgment. If the models need to learn from the real world, that could also limit the reach of self-improvement.
然后是训练数据的问题。AI 最近的许多进展都是在模型可以通过“可验证的奖励”自学如何成功的领域。一段软件要么运行，要么不运行；数学证明要么正确，要么不正确。在这种情况下，纯粹为训练其他模型而由模型生成的合成数据可以检查准确性并添加到训练数据中，而不会冒用 AI 自己的输出训练时通常伴随的退化风险。让模型在创意写作或法律判断方面变得更好则更为棘手。如果模型需要从现实世界中学习，这也可能限制自我改进的范围。
第十九段
“Closing the loop” may be a step on the road to superintelligence and—depending on your disposition—utopia or doom. But it is not the only step required to produce exponential growth in AI’s capabilities. ■
“闭环”可能是通向超级智能的道路上的一步——取决于你的倾向——乌托邦或末日。但这并不是实现 AI 能力指数级增长所需的唯一步骤。■
