递归式自我升级：AI的进化之路与人类的选择

"递归式自我升级"对这项技术究竟意味着什么？它将把AI带向何方？

Anthropic即将IPO，估值有望创下历史纪录——其代码生成助手Claude Code已贡献了公司80%的产出，也成为全球开发者必备工具。然而，正值鼎盛时期的Anthropic却在6月5日公开呼吁世人"有权选择放慢甚至暂停前沿AI开发"。这种看似"自断经脉"的倡议背后，隐藏着一种被称为"递归式自我升级"（RSI）的可能性：模型能够创造出比自身更强、更快的下一代，下一代再继续创造下一代，如此循环往复。一旦这一闭环形成，AI便可脱离人类独立迭代，"奇点"或"快速起飞"（学界戏称为"砰然一声foom"）的预言将不再是科幻。但实现全流程RSI的路上，仍横亘着几道坎：算力瓶颈、训练数据匮乏、安全验证仍需人类把关。CSET警告称，AI主导的研发可将生产力提升千倍，但也意味着人类或将在"自己设定目标、自己训练模型、自己检验安全"的循环中彻底失去控制权。

原文 When ANTHROPIC, an artificial-intelligence lab,debutson stock markets later this year, it is likely to be one of the biggestinitial public offeringsin history. That's because Claude, the company's chatbot isbeloved ofcoders, who are willing to pay a lot for access. Since Claude Code, its software-engineering agent, launched in February 2025, it has becomeindispensablefor developers around the world. That includes Anthropic's own: more than four-fifths of the code it published in May was written by Claude, the company says. Before Claude Code, the percentage was "low single-digits".

翻译 人工智能实验室Anthropic将于今年晚些时候上市，很可能会成为史上规模最大的IPO之一。原因无他：该公司旗下的聊天机器人Claude深受程序员青睐，大家心甘情愿为此付费。2025年2月，编程助手Claude Code一经发布，便迅速成为全球开发者离不开的工具。Anthropic自身也不例外——公司表示，5月发布的代码中，超过八成出自Claude之手；而在Claude Code出现之前，这一比例还只是可怜的个位数。

原文 The systems have improvedin quality ofoutput as well as quantity. An influentialbenchmarkfrom METR, a think-tank, shows that in early 2025 Anthropic's models could complete tasks that took human engineers a little under an hour. The company's latest systems can complete tasks that would take more than a working day.

翻译 这些系统的进步体现在质与量两方面。智库METR的一项权威基准测试显示：2025年初，Anthropic的模型仅能处理人类工程师耗时不到一小时的任务；如今，其最新系统已能完成需要一个完整工作日才能搞定的工作。

原文 And so it may be easy to raise acynicaleyebrowwhen the company,at the top of its gameandoutclassingthe competition, calls for the world to have "the option to slow or temporarily pausefrontierAI development", as it did on June 5th. What market leader would not wish that its competition stop trying to catch up?

翻译 因此，当这家正值巅峰、已将竞争对手远远甩在身后的公司，于6月5日呼吁世人"有权选择放缓乃至暂停前沿AI开发"时，旁观者难免心生疑虑：市场领头羊又岂会真心希望追赶者停下脚步？

原文 I, robot. Yet Anthropic's leaders, who have for years worried about the prospect of out-of-control AIwreakinghavoc, seemsincere. The latest generation of AI models are suchcompetentcoders, engineers and (soon) scientists that many worry they may be among the last ever made by humans. Jack Clark, an Anthropic co-founder, thinks there is a 60% chance that, by the end of 2028, an AI system will be capable of creating its ownsuccessorwith no humaninvolvementat all.

翻译 【我，机器人】然而，Anthropic的领导者们多年来始终担忧AI失控可能带来的灾难，此番呼吁似乎出于真心。最新一代AI模型在编程、工程乃至（即将涉及的）科研领域都已游刃有余，以至于许多人隐隐担忧：人类亲手打造的最后一批AI，或许就是眼前这几代。联合创始人Jack Clark认为，到2028年底，AI系统有六成概率能在完全无需人类参与的情况下，自主创造出"接班人"。

原文 That moment would mark the beginning of a process called "recursive self-improvement" (RSI), a closed loop. Version one of a model produces version two, which is faster and more capable; version two produces version three, which is more so again. The loop continues, and the improvements grow with eachiteration. Build an AI system capable of this, and your human engineers never need to build another one again. "What can seem to many like afancifulstory may instead be a real trend," says Mr Clark.

翻译 那一刻，将标志着一个名为"递归式自我升级"（RSI）进程的开启——一个首尾相连的闭环。第一代模型孕育出第二代，速度更快、能力更强；第二代再孕育出第三代，更胜一筹。循环持续，每一轮迭代都带来更显著的提升。一旦这样的系统成真，人类工程师便再无需亲手打造下一代。"许多人眼中的天方夜谭，或许正悄然演变为真实趋势。"Clark如是说。

原文 Nobody knows for sure what the consequences of recursive self-improvement would be. Because AI can, unlike humans, work tirelessly and constantly, some think it wouldin short orderlead to asuperintelligentAI—a "fast take-off". (It has also beenonomatopoeicallydubbed "going foom", for the sound one might imagine an intelligence explosion making). AIdoomersfear the superintelligence would be beyond human control, and that the start of RSI is the moment at which humanity's fate ishanded over tothe machines. Yet a self-improving AI would probably face speed limits, at least at first.

翻译 递归式自我升级究竟会带来什么后果，无人确知。AI与人类不同，它能不知疲倦、持续运转——有人据此认为，超级智能的"快速起飞"将指日可待。（业界甚至用拟声词称之为"going foom"——仿佛智能爆炸时那声轰然巨响。）"末日论者"担心超级智能一旦失控，人类便无力驾驭；而RSI的起点，就是人类将命运交给机器的那一刻。然而，至少在初期，AI的自我升级恐怕仍会遭遇"限速"。

原文 Building a model capable of RSI would require automating a range of specialist tasks currently carried out by humans. At present data scientists work on the theory of AI and coders put it into practice. Systems engineers build the foundations on which toy models can be raised to production scale. Other people seek out novel sources of training data, or experiment with ways to generate it fresh. Alignment and safety teams check that what comes out of the training process won't cause harm,intentionalor otherwise.

翻译 要打造真正具备RSI能力的模型，就必须将目前由人类承担的一系列专业任务全部自动化。目前，数据科学家负责钻研AI理论，程序员负责代码实现；系统工程师构建底层基础设施，使实验模型能够扩展到生产级别；还有人寻找新的训练数据来源，或尝试从头合成；对齐与安全团队则把守最后一关——确保训练出的模型，无论有意无意，都不会造成危害。

原文 The joy of repetition. Not all of those teams are equallyamenableto AI assistance, and within each specialism some tasks are moreautomatablethan others. It will not be too long until a human coder can do their job without ever writing a line of computer code themselves, but it may be some time until an AI is able tonegotiateto acquire a previouslyundigitisedcollection of scientific papers.

翻译 【重复的力量】这些团队对AI辅助的接受程度并不一致，即便在同一专业领域内，各项任务的自动化潜力也参差不齐。程序员无需亲笔写下一行代码就能完成工作的日子已为期不远；但让AI亲自出面谈判收购一套从未数字化的科学文献库，恐怕还为时尚早。

原文 It is not always obvious how the "jagged frontier" will progress. Designing new algorithms seemed one of the safer jobs, until one of Google DeepMind's models, AlphaEvolve, began doing it in May 2025. It proposed a change to how Google spreads workloads across its data centres that saved 0.7% of the company's worldwide computing power, and found better ways to perform matrix multiplication, which speeded up the training of Gemini, the company'sflagshiplarge language model (LLM), by 1%.

翻译 这条"参差不齐的前沿"将向何处延伸，历来难以预料。设计新算法曾被视为稳妥的工作，直到2025年5月Google DeepMind的AlphaEvolve模型也开始涉足这一领域。它提出了一项改进谷歌数据中心负载分配的方案，节省了该公司全球计算资源的0.7%；还找到了更优的矩阵乘法方法，使旗舰大模型Gemini的训练速度提升了1%。

原文 Full RSI requires every task in this chain to become automated. The AI-powered acceleration of research and development (R&D) may be felt before then, however. "As the fraction of AI R&D performed by AI systems increases, the productivity boost over human-only R&D" could increase ten-fold, then a hundred-fold, then a thousand-fold, according to a report published in January by the Centre for Security and Emerging Technology (CSET), a think-tank within Georgetown University. In that scenario, it warns that even if some aspects of AI R&D are initially difficult to automate, "the accelerated rate of progress means thosebottlenecksare soon overcome."

翻译 完全意义上的RSI要求链条上每个环节都实现自动化。然而在此之前，AI对研发的加速效应或许早已显现。乔治城大学旗下智库CSET今年1月发布的报告指出："随着AI系统中由AI完成的研发占比不断攀升，相较于纯人类研发，其生产力优势将从十倍跃升至百倍、千倍。"报告同时警告：即便AI研发的某些环节起初难以自动化，"加速的进步速度意味着这些瓶颈将很快被突破。"

原文 Today no AI model can build its ownsuccessor. But big AI models can build smaller models on their own. With human help they can build other big AI models, too. Earlier this year Andrej Karpathy, a then-independent researcher who now works for Anthropic, trained a chatbot about as capable as GPT-2, a large language model built by OpenAI in 2019. Back then the model took 168 hours of training to build on 32 state-of-the-art chips; Dr Karpathy achieved the same result using a single computer with eight GPUs, the specialised chips used to build AI, in only three hours. With some more months of work he reduced the training time for his model, Nanochat, to just over two hours.

翻译 目前，尚无AI模型能独立打造自己的"继任者"。但大模型自行构建小模型已不在话下；若有人类协助，再造一个大模型也轻而易举。今年早些时候，彼时仍是独立研究员、现就职于Anthropic的Andrej Karpathy，训练出一款能力与GPT-2相当的聊天机器人。GPT-2是OpenAI于2019年推出的大语言模型，当年需要32块顶级芯片、耗时168小时才能训练完成；而Karpathy仅用一台配备8块GPU（AI训练专用芯片）的普通电脑，3小时便复现了同等效果。此后又经过数月打磨，将其模型Nanochat的训练时间进一步压缩到两个多小时。

原文 In March he handed the work of speeding up the training process over to an AI agent called AutoResearch. In two days the training time dropped to one hour and 48 minutes, and five days after that it fell to one hour and 39 minutes. "I didn't touch anything," Dr Karpathy says. The 18% improvement on the human work isstrikingbecause Dr Karpathy is a particularly talented human: he was a founding member of the research team at OpenAI and the head of AI at Tesla for five years.

翻译 今年3月，他将进一步提速的任务交给了一个名为AutoResearch的AI代理。两天内，训练时长从两个多小时降至1小时48分；又过了五天，再降至1小时39分。"我全程没有碰过一下。"Karpathy表示。能在人类成果基础上再提升18%，这事本身就够惊人——毕竟Karpathy绝非等闲之辈，他是OpenAI研究团队的创始成员，还曾在特斯拉执掌AI部门长达五年。

原文 The improvements themselves wereprosaic. The AI agent picked better starting values for the training run, widened the scope of the LLM's "attention" window and noticed that the model's focus was wandering. None of this is particularlynovel, Dr Karpathy says. But he had missed them. "Theystack upand actually improved Nanochat," he says.

翻译 改进本身其实平平无奇：AI代理为训练选择了更优的初始值，拓宽了LLM的"注意力"窗口，还发现模型的注意力有所分散。Karpathy说，这些都算不上什么新花样。但偏偏他自己之前没注意到。"它们累积起来，确实让Nanochat得到了提升。"他说。

原文 Speed-ups of this kind areinevitableas models become more capable. Much of the work of buildingterabyte-sizefrontier models is lessglamorousthan the AI industry'senormoussalaries and fancy offices suggest. It involvesplumbingtogether the layers of an infrastructure stack that are bought in from third parties, debugging hardware and software set-ups andtweaking"hyperparameters", the initial set-up of a training run, until the outcome looks solid. An AI system can do much of that today, with littlesupervision.

翻译 随着模型能力日益强大，此类提速已是大势所趋。然而，构建TB级别前沿模型的大量幕后工作，远非AI行业的天价薪酬和豪华办公室所呈现的那般光鲜：不过是将采自第三方的层层基础设施拼接整合，调试软硬件配置，反复调整"超参数"——即训练启动前的各项初始设置——直至结果稳定可靠。如今，AI系统已能在极少监督下独立完成这些工作。

原文 But even the morenuancedintellectual work is nearing automation, says Joe Spisak, a researcher at Reflection AI, a lab based in New York that is building frontier models that areopen-weight(meaning their parameters are publicly released). Give a frontier system a rough sketch of an idea for efficiency gains, and it is increasingly capable of designing an experiment, running tests on a toy model, seeing what works and responding with a plan that is ready to implement at scale.

翻译 然而，就连那些更为精妙的智力工作也正逐步走向自动化。Reflection AI是纽约的一家实验室，正在研发开放权重的前沿模型（即模型参数公开可查）。该实验室研究员Joe Spisak表示，只需给前沿系统一个效率提升的大致思路，它便已越来越能自行设计实验、在玩具模型上运行测试、观察哪些方法奏效，并输出一份可直接大规模实施的方案。

原文 AI models can carry out these sorts of tasks, which take humans hours, in around 30 minutes. Increasingly, humans play the role only of research director,steeringthe AI to run experiments, which the models code up, debug, optimise and monitor themselves. The productivity boost isalluring, but also alarming. As the role that humans play in the production processshrinks, they may lose control. The end result could be models trained by models, to achieve goals set by models, whose safety is verified only by models.

翻译 这些原本需要人类耗费数小时的任务，AI模型大约三十分钟就能搞定。如今，人类越来越只扮演"研究总监"的角色——引导AI执行实验；至于代码编写、调试排错、性能优化与全程监控，则统统由模型自行完成。效率的跃升固然诱人，却也令人不安：随着人类在生产流程中的角色日渐萎缩，失控的风险也在攀升。最终可能是模型训练模型、模型设定目标、模型验证安全——整条链条中，人类的痕迹几乎消失殆尽。

原文 Some fear a disaster. Max Tegmark, a physicist and machine-learning researcher at the Massachusetts Institute of Technology who hasdevotedmuch of the past decade tocampaigningfor AI safety,likensit to a driverflooring the acceleratoron themotorwaywith their eyes closed. The result would be certaindoom, he told the The Economist's "Inside Tech" video show, as long as the driver refuses to open their eyes. Powerful AI systems couldoutcompetehumans as the decision makers in government and commerce, says Professor Tegmark,disempoweringhumanity; they could offer supreme power to whoever first builds them,ushering inglobaltotalitarianism; or they could simplyceaseto care about humanity at all, and gradually squeeze people out to make room for more data centres and power generation.

翻译 有人看到的却是灾难。麻省理工学院物理学家、机器学习研究者Max Tegmark，过去十年不遗余力地为AI安全奔走呼吁。他将此比作司机蒙着眼在高速公路上油门踩到底。他在《经济学人》"Inside Tech"视频节目中表示，只要这位司机始终不肯睁眼，结局注定是在劫难逃。Tegmark教授认为，强大的AI系统可能在政府与商业决策中全面超越人类，令人类丧失话语权；也可能将无上权力拱手交给最先造出它的人，催生全球极权统治；又或者，它干脆彻底漠视人类存在，逐步将人排挤出去，腾出空间建造更多数据中心和发电设施。

原文 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. Thiswarrantspreparatoryaction now." Anthropic, it seems, is close to agreeing with that idea.

翻译 三年前，Tegmark教授曾领衔呼吁全球暂停AI开发，理由是彼时最先进的GPT-4问世，如同那场蒙眼狂奔。而今年CSET的报告则警告：由RSI创造的系统"蕴含极端风险，亟需现在就采取预防措施。"Anthropic似乎也在向这一观点靠拢。

原文 Hot chip. 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 theirpredecessors, forcing progress to occur at the pace of data-centre development.

翻译 【发烫的芯片】此外，还有几道物理限制，短期内仍将制约模型自我升级的速度。最关键的便是算力。尽管效率已有提升，新一代模型训练所需的算力仍较前辈有增无减，进度不得不受制于数据中心的建设速度。

原文 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 outopen-endedR&D. The more demand there is in the first category, the less capacity, in the short term, there is for the other two.

翻译 CSET临时执行所长、上述报告首席作者之一Helen Toner指出，消费者对AI的使用需求也将拖累AI驱动的研发。数据中心算力本就紧张，须在服务付费用户、训练未来模型、支撑开放式研发三者之间谨慎分配。短期内，第一类需求越大，后两类获得的算力就越少。

原文 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 casessynthetic data, generated by models purely to train other models, can be checked for accuracy and added to the training data without risking thedegeneracythat 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 thereachof self-improvement.

翻译 再就是训练数据的问题。AI近期的长足进步大多来自"可验证奖励"机制发挥作用的领域：代码能否运行、数学证明对错与否，都有硬性标准。在这类场景中，由模型专为训练其他模型而生成的"合成数据"可被精准核验后加入训练集，不必担心"自我投喂"导致的退化问题。但要提升模型的创意写作或法律判断能力，就没那么简单了。一旦模型需要从真实世界汲取养分，"自我升级"的空间便会受到限制。

原文 "Closing the loop" may be a step on the road to superintelligence and—depending on your disposition—utopiaordoom. But it is not the only step required to produce exponential growth in AI's capabilities.

翻译 "合上闭环"或许是通往超级智能之路上的一个台阶，其终点——取决于你的心态——或是乌托邦，或是末日。但它绝非AI能力指数级增长所需的唯一一步。

原文：WHEN ANTHROPIC, an artificial-intelligence lab, debuts on stock markets later this year, it is likely to be one of the biggest initial public offerings in history. 翻译：人工智能实验室Anthropic将于今年晚些时候上市，很可能成为史上规模最大的IPO之一。 解析：

原文：Not all of those teams are equally amenable to AI assistance, and within each specialism some tasks are more automatable than others. 翻译：这些团队对AI辅助的接受程度并不一致，即便在同一专业领域内，各项任务的自动化潜力也参差不齐。 解析：

原文：It is not always obvious how the "jagged frontier" will progress. 翻译：这条"参差不齐的前沿"将如何延伸，目前尚难预料。 解析：

原文："As the fraction of AI R&D performed by AI systems increases, the productivity boost over human-only R&D" could increase ten-fold, then a hundred-fold, then a thousand-fold, 翻译："随着AI系统中由AI完成的研发占比不断攀升，相较于纯人类研发的生产力优势将从十倍跃升至百倍、千倍。" 解析：

原文：Earlier this year Andrej Karpathy, a then-independent researcher who now works for Anthropic, trained a chatbot about as capable as GPT-2, a large language model built by OpenAI in 2019. 翻译：今年早些时候，彼时仍是独立研究员的Andrej Karpathy（现就职于Anthropic）训练出一款能力与GPT-2相当的聊天机器人——后者是OpenAI于2019年推出的大语言模型。 解析：

原文："I didn't touch anything," Dr Karpathy says. The 18% improvement on the human work is striking because Dr Karpathy is a particularly talented human: he was a founding member of the research team at OpenAI and the head of AI at Tesla for five years. 翻译："我一个指头都没动过。"Karpathy博士如是说。这18%的提升之所以令人侧目，是因为Karpathy本人绝非等闲之辈——他既是OpenAI研究团队的创始成员，又曾执掌特斯拉AI部门五年。 解析：

原文：The improvements themselves were prosaic. The AI agent picked better starting values for the training run, widened the scope of the LLM's "attention" window and noticed that the model's focus was wandering. None of this is particularly novel, Dr Karpathy says. But he had missed them. 翻译：改进的内容本身其实平平无奇：AI代理为训练挑选了更优的初始值，扩宽了大语言模型的"注意力"窗口，还察觉到模型注意力有"走神"之嫌。Karpathy博士坦言，这些点子并无石破天惊之处，只是他自己先前未曾察觉。 解析：

原文：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. 翻译：麻省理工学院物理学家、机器学习研究者Max Tegmark过去十年不遗余力地奔走呼吁AI安全，他将此情景比作"蒙眼司机在高速公路上一脚油门踩到底"。 解析：

原文：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. 翻译：他在《经济学人》"Inside Tech"视频节目中说道：强大的AI系统可能在政府与商业决策上全面超越人类，令其"被剥夺权力"；也可能将至高无上的权力拱手送给率先造出它的国家，催生全球极权；甚至干脆对人类置之不理，逐步将人挤到一边，只为腾出更多空间来兴建数据中心与发电厂。 解析：

原文：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. 翻译：CSET临时执行所长、上述报告首席作者之一Helen Toner指出，消费者对AI的使用需求，也将掣肘AI研发本身。 解析：

原文：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. 翻译：数据中心算力本就捉襟见肘，须在服务付费用户、训练未来模型、支撑开放式研发三者之间精打细算。 解析：

原文："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. 翻译："合上闭环"或许是通往超级智能征途上的一个台阶，其终点——取决于你的性情——或是乌托邦，或是末世。但它绝非AI能力指数级增长所需的唯一台阶。 解析：