2025年的人工智能：整体

bytechnicalities作者：technicalities8th Dec 2025 This is the editorial for this year’s "Shallow Review of AI Safety". (It got long enoughto stand alone.) 这是今年“AI安全浅层综述”的社论。（它写得⾜够长，可以独⽴成⽂。） Epistemic status: subjective impressions plus one new graph plus 300 links. 认知状态：主观印象，加上⼀张新图表和300个链接。 Huge thanks to Jaeho Lee, Jaime Sevilla, and Lexin Zhou for running lots of tests pro bonoand so greatly improving the main analysis. ⾮常感谢Jaeho Lee、Jaime Sevilla和Lexin Zhou⽆偿进⾏了⼤量测试，从⽽⼤⼤改进了主要分析。 tl;dr要点概述 Informed peopledisagreeabout the prospects for LLM AGI – or even just whatexactly was achieved this year. But the famous ones with abookto talk at leastagree that we’re2-20years off (allowing for other paradigms arising). In thispiece I stick to arguments rather than reporting who thinks what. 有见识的⼈对LLM AGI的前景持有不同意见——甚⾄对今年究竟实现了什么也有分歧。但那些出书、出⾯发声的名⼈⾄少⼀致认为我们还差2到20年（并且允许出现其他范式）。在这篇⽂章中，我坚持论证⽽不是报道谁持何种观点。微信公众号 404K微信公众号 404K My view: compared to last year, AI is much more impressive but notproportionally more useful. They improved on some things they were explicitlyoptimised for (coding, vision, OCR, benchmarks), and did not hugely improveon everything else. Progress is thus (still!) consistent with current frontier 总体上我真希望能告诉你⼀个数字——预期的净安全变化（今年危险能⼒与代理性能的提升，减去能促进对齐的那部分能⼒提升，减去迄今为⽌实际实施的最佳对齐与控制技术组合的累积效应）。但我做不到。 Capabilities in 20252025年的能⼒ Better, but how much?更好了，但是多少？ Arguments against 2025 capabilities growth being above-trend反对2025年能⼒增长超出趋势的论点 So pretraining can't scale yet, because most inference chips aren't big enough tohandle trillions of active parameters. And scaling RL more won't help as muchas it helped this year, because of inefficiency. So... 所以预训练还不能扩展，因为⼤多数推理芯⽚不够强⼤，⽆法处理数万亿个活动参数。再多扩展强化学习也不会像今年那样带来那么⼤的帮助，因为效率问题。所以…… By late 2025, the obsolete modal “AI 2027” scenario described the beginning ofa divergence between the lead lab and the runner-up frontier labs.This isbecause the leader’s superior ability to generate or acquire new training dataand algorithm ideas was supposed to compound and widen their lead. Instead,we see the erstwhile leader OpenAI and some others clustering around thesame level, which is weak evidence that synthetic data and AI-AI R&D aren’tthere yet. Anthropic are makinglarge claimsabout Opus 4.5’s capabilities,somaybethis will arrive on time next year.[4] 到2025年末，过时的“AI 2027”情景描述了领先实验室与紧随其后的前沿实验室之间开始出现分化的情形。这是因为原先认为领先者在⽣成或获取新的训练数据和算法想法⽅⾯具有更强的能⼒，这种能⼒会复利并扩⼤他们的领先优势。然⽽，我们看到曾经的领导者OpenAI和其他⼀些组织聚集在相同的⽔平上，这在弱证据上表明合成数据和AI对AI的研发还未成熟。Anthropic正在对Opus 4.5的能⼒做出很⼤声明，所以也许这会在明年按时到来。[4]微信公众号 404K微信公众号 404K For thefirst time there are nowmanyexamples of LLMs helping with actualresearch mathematics. But if youlook closelyit’s all still in-distribution in thebroad sense: new implications of existing facts and techniques. (I don’t mean todemean this; probably most mathematicsfits this spec.) And it's almost neverfully autonomous; there's usually hundreds of bits of human steering. 现在⾸次出现了许多LLMs在实际研究数学中提供帮助的例⼦。但如果你仔细观察，从⼴义上看这仍然是在分布内：对已有事实和技术的新推论。（我并不是要贬低这点；⼤多数数学可能都符合这个定义。）⽽且⼏乎从未完全⾃主；通常有数百处⼈类的引导。 Extremelymixedevidenceon the trend in the hallucination rate.关于幻觉率趋势的证据极为混杂。 Companies make claims about their one-million- or ten-million-tokeneffectivecontext windows,butIdon’tbelieve it. 公司声称其有效上下⽂窗⼜达到⼀百万或⼀千万标记，但我不相信这些说法。 In lieu of trying the agents for serious work yourself, you could at least look atthehighlightsof thegullibleand precompetent AIs in theAI Village. 如果你不准备亲⾃将这些代理⽤于严肃⼯作，⾄少可以看看AI Village中那些轻信且能⼒不⾜的AI的精彩⽚段。微信公众号 404K微信公众号 404K Arguments for 2025 capabilities growth being above-trend 关于2025年能⼒增长超出长期趋势的论据 We now have measures which are a bit more like AGI metrics than dumb single-taskstatic benchmarks are. What do they say? 我们现在有⼀些更接近通⽤⼈⼯智能（AGI）指标的测量⽅法，⽽不是那些愚蠢的单任务静态基准。它们显⽰了什么？ Ignoring the (nonrobust)ECI GPT-2 rate, we can sayyes: 2025 is fast, as fast asever, or more.[6] 忽略（不稳健的）ECI GPT-2速率，我们可以说是的：2025年很快，和以往⼀样快，甚⾄更快。[6] Even though these are the best we have, we can’t defer to these numbers.Whatelse is there?[7] 尽管这些是我们现有的最好结果，但我们不能把这些数字当作最终结论。还有什么其他可能性吗？[7] In May they passed some threshold and Ifinally started using LLMs for actualtasks. For me this is mostly due to the search agents replacing a degradedGoogle search. I‘mnottheonly onewhoflipped this year. This hasty poll isworth more to me than any benchmark: 在五⽉份，他们通过了某个阈值，我终于开始在实际任务中使⽤LLMs。对我来说，这主要是因为搜索代理取代了变差的Google搜索。我并不是今年才改变看法的唯⼀⼀个⼈。这个匆忙的民意调查对我来说⽐任何基准测试都更有价值： Or if you prefer aformal study(n=2,430 researchers):或者如果你更喜欢⼀项正式研究（样本量= 2,430名研究⼈员）： On actual adoption and real-world automation: 关于实际采⽤和现实世界⾃动化： Based on self-reports, theSt Louis Fedthinks that “Between 1 and 7% of allwork hours are currently assisted by generative AI, and respondentsreport time savings equivalent to 1.4% of total work hours… across allworkers (including non-users… Our estimated aggregate productivitygain from genAI (1.2%)”. That’s model-based, using year-old data, andnaively assuming that the AI outputs are of equal quality. Not strong.微信公众