字号·· | 护眼
经济学人

预计人工智能技术将迎来加速发展的一年Expect a year of accelerating technological progress in AI

点「原文对照」整页切到原文,或双击某段只看那段的原文。

多年来,人们对人工智能做出过许多听起来古怪、疯狂的预测。它将解决气候变化、治愈所有疾病、让人类长生不老!它将把整个世界变成回形针!它将变得如此先进,以至于把人类当宠物养!人们很容易耸耸肩或翻个白眼,就把这种推测一笔带过。但到了2026年,人工智能真的开始做一些听起来古怪、疯狂的事情了。

FOR YEARS, people have made strange, crazy-sounding predictions about artificial intelligence. It will solve climate change, cure all disease and make people immortal! It will turn the whole world into paper clips! It will be so advanced that it will treat humans like pets! It was easy to dismiss such speculation with a shrug or an eye-roll. But in 2026 AI actually started to do strange, crazy-sounding things.

在几起独立事件中,成群的AI智能体失控,冲出AI实验室并入侵外部系统。加密钱包被掏空,欧洲各地超市的冰箱被设置为除霜模式,这一切都发生在AI辅助的网络攻击中。AI模型给出了解决数学、计算机科学和流体力学长期难题的方案,引发学术界大量灵魂拷问。AI撰写的故事获得文学奖。AI生成的音乐登上排行榜榜首。

In several separate incidents, swarms of AI agents went rogue, breaking out of AI labs and breaching external systems. Crypto wallets were drained, and supermarket fridges across Europe set to defrost, in AI-assisted cyber-attacks. AI models produced solutions to long-standing problems in maths, computer science and fluid dynamics, causing much soul-searching in academia. AI-written stories won literary prizes. AI-generated music topped the charts.

突然间,AI进步可能导致意想不到且往往令人恐惧的结果这一风险,似乎变得更可信了。事情看起来可能已经够奇怪了。但在2027年,它们将变得更加奇怪。

Suddenly, the risk that progress in AI could lead to unexpected and often frightening outcomes seems much more plausible. Things may seem strange enough already. But in 2027 they will get even stranger.

原因之一是,AI在编程、黑客攻击和数学方面已经具备超人类水平,可能即将变得强大得多。业内许多人预期,2027年将迎来“递归自我改进”(RSI)的到来:即AI系统以有意义的方式改进其自身的继任者,形成反馈循环,从而导致AI能力爆发式增长。一些人担心,跨越这一门槛将导致人类失去对该技术的控制,因为它开始以人类无法理解的方式自我改进。而RSI可能在科学领域引发混乱,因为AI模型推导出令人类困惑的新结果,这种情况在数学领域已经出现。

One reason is that AI, already superhuman at coding, hacking and maths, may be about to become far more powerful. Many in the industry expect 2027 to herald the arrival of “recursive self-improvement” (RSI): the process whereby AI systems meaningfully improve their own successors, in a feedback loop that leads to an explosion in AI capabilities. Some worry that crossing this threshold will cause mankind to lose control of the technology, as it starts to improve itself in ways humans cannot comprehend. And RSI could cause havoc in science, as AI models derive new results that baffle humans, as is already happening in maths.

在网络安全领域,失控可能已经发生。2026年初,Anthropic表示其最强大的内部模型“Mythos”过于危险,不宜公开发布。随着2027年的到来,这些能力已扩散开来。中国实验室Moonshot(月之暗面)的开源模型Kimi K3能够发现并利用网络安全漏洞,而这类能力曾是美国国家级网络力量的专属,且几乎无法控制谁能获得这种力量。在2026年的一次AI大会上,大多数行业领袖表示,他们预计未来几年内将发生一起造成1000亿美元损失或1万人伤亡的网络安全事件。

When it comes to cyber-security, the loss of control may already have occurred. In early 2026 Anthropic said its most powerful internal model, Mythos, was too dangerous to release openly. As 2027 dawns, those capabilities have proliferated. Open-source models such as Kimi K3 from Moonshot, a Chinese lab, are capable of finding and exploiting cyber-security flaws that were once the preserve of America’s national champions, with little ability to control who has access to that power. At an AI conference in 2026, a majority of leaders in the industry said they expected a cyber-security incident costing $100bn, or 10,000 lives, to happen in the next few years.

理论上,AI将改善网络安全,让工程师能在软件发布前分析代码、发现并修补漏洞。但实际上,世界上已存在大量老旧、脆弱的软件。供应商正争相发布安全更新:微软过去每月发布70至80个补丁,但在9月发布了972个,10月发布了1411个。然而,并非所有软件都在积极维护,也并非所有补丁都会被应用(想想那些超市冰柜)。因此,未来几个月网络攻击的浪潮将会上涨。

In theory, AI will improve cyber-security, by allowing engineers to analyse software, and find and patch flaws, before it is released. In practice, there is a lot of old, vulnerable software out in the world already. Vendors are rushing to issue security updates: Microsoft, which used to issue 70-80 patches per month, issued 972 in September and 1,411 in October. But not all software is actively maintained, and not all patches are applied (think of all those supermarket fridges). So the tide of cyber-attacks will rise in the coming months.

在其他方面,AI行业也面临着一个充满考验和动荡的一年。针对数据中心的舆论和监管反弹正在加剧,威胁到新基础设施的建设,令投资者感到不安。Anthropic的首次公开募股将使其财务接受审视;OpenAI预计将跟进。实验室面临着两条战线的财务压力。

The AI industry faces a testing and volatile year in other respects, too. A popular and regulatory backlash against data centres is intensifying, threatening the construction of new infrastructure and spooking investors. Anthropic’s public offering will expose its finances to scrutiny; OpenAI is expected to follow. And the labs face financial pressure on two fronts.

首先,研发开支没有显示出放缓的迹象。训练运行的巨大成本正因昂贵的“后训练”而加剧,人工智能模型的能力正是在这一阶段得到打磨。如果实现规模定律(RSI),这些能力将会飙升,但成本也会随之激增:人工智能软件工程师在计算能力上的花费速度,远比任何人类都要快。难怪根据Anthropic公司10月发布的招股书,该公司预计在未来几年内将花费7500亿美元用于计算能力。

First, spending on R&D shows no sign of slowing. The huge cost of training runs is being exacerbated by expensive “post-training” in which AI models’ abilities are honed. If RSI kicks in, those abilities will surge, but so will costs: AI software engineers can spend money on computing power far more quickly than any human. No wonder that, according to Anthropic’s prospectus, released in October, the company expects to spend $750bn on computing power in the coming years.

与此同时,美国各大实验室正面临越来越大的定价压力。来自中国人工智能实验室的竞争迫使它们削减价格,而这场竞争远未结束。10月,中国实验室深度求索(DeepSeek)将其模型的运行软件在网上免费提供,以帮助用户在自己的硬件上运行人工智能模型,而不是支付基于云的服务费用。在台式机上运行人工智能模型目前主要还是一种业余爱好。

At the same time, the big American labs face growing pricing pressure. Competition from Chinese AI labs has forced them to cut their prices, and the fight is far from over. In October DeepSeek, a Chinese lab, made the software for running its models available free online, to help users run AI models on their own hardware rather than paying for cloud-based services. Running AI models on desktop machines is largely a hobby for now.

但斯坦福大学的研究人员发现,对于80%以上的任务,小型人工智能模型与基于云的模型表现得一样好。对许多用户来说,一两年内在笔记本电脑或智能手机上运行的免费本地模型可能就足够好了——这对美国各大实验室来说将是一个噩梦般的场景。

But researchers at Stanford University found that small AI models were just as good as cloud-based models for over 80% of tasks. For many users, a free, local model running on their laptop or smartphone may be good enough within a year or two—a nightmare scenario for the big American labs.

但这对于人工智能用户来说将是好消息。随着竞争压力压低成本,并且在预先承诺的投资推动下能力持续提升,其结果可能类似于千禧年之交电信繁荣所推动的连接大爆发。当然,那场繁荣对电信公司本身并没有以良好的结局收场。过去五年人工智能行业的故事一直是期望值和估值飙升,而现实却在苦苦追赶。准备迎接2027年截然相反的情况:随着人工智能系统开始以日益奇特的方式改变世界,集体的不相信可能会画上句号。

But it would be good news for users of AI. With costs pushed down by competitive pressure, and capabilities continuing to improve thanks to pre-committed investment, the outcome could resemble the explosion in connectivity powered by the telecoms boom at the turn of the millennium. That boom did not, of course, end well for the telecoms companies themselves. The story of the past five years of the AI industry has been of expectations and valuations soaring, as reality struggled to keep up. In 2027 prepare for the opposite: as AI systems start to change the world, in increasingly strange ways, the collective suspension of disbelief may come to an end. ■

阅读媒体原文 ↗