Howie和小能熊

什么是”世界模型“?它为何重要?

人类的大脑,

本质上就是一个终身自我训练的世界模型。

什么是世界模型?它为什么重要?

从sora开始,“世界模型”这个词就走进了我们的日常。 那么,sora里真的有世界模型吗?最新最强的llm,例如gpt-5,有世界模型吗?

hinton、ilya、demis hassabis、李飞飞和杨立昆每个人对“如何实现世界模型”的观点都不一样,到底谁说的对?

可以这么理解:所谓世界模型,就是就是世界在人脑或ai的神经网络中的内部表征(心理模型)。在“大脑”里对外部世界形成抽象表示,形成内部模拟,这可以用于理解、预测和规划外部环境的交互。

我们的大脑并不是一台录像机,并不存储海量的原始感官数据;相反,大脑构建了一个关于世界如何运作的、经过压缩的、可预测的内部模型。人知道苹果熟了会掉下来而不是飞上去,sora等视频生成ai知道人走在水面会沉下去而不是如履平地,这都是“世界模型”的作用。

虽然围绕“世界模型”还有很多争辩(想一想杨立昆),但这个概念,不论是哲学层面的、还是认知心理学层面、ai技术层面,都对普通人是有意义且有启发的:

人类的大脑,本质上就是一个终身自我训练的世界模型。终身学习者不是要成为信息的收集者,而是要成为世界模型的构建者;

学习的目标是要在自己关心的领域内,建立一个融会贯通的“心理模型”体系。学习是构建模型,而不是堆砌信息;真正的聪明,不在于记住了多少知识,而在于用于多好的世界模型(解释世界、解决问题);

最好的学习方法,一定是主动预测的,而非被动接受的。人脑一直在做预测,只有在不断的预测并与现实比对、修正的过程中,模型才会变得更强大。不论是读书还是课程学习,主动预测都胜过被动灌输。这种主动的“构建模型——预测——反馈迭代”循环,是最高效的学习方式。

最近读到一篇五星级好文,来自最优秀的科学杂志之一《量子杂志》(quanta magazine),把“世界模型”这个基础概念的来龙去脉讲得很清楚。所以,我特别制作了双语对照版本,分享给你。祝阅读愉快~ ❤️

“世界模型”,

ai领域的一个古老概念焕发新生

You’re carrying around in your head a model of how the world works. Will AI systems need to do the same?

你的大脑里装着一个关于“世界如何运转”的心理模型。AI 系统是否也需要拥有同样的模型?

Image

title: ‘World Models,’ an Old Idea in AI, Mount a Comeback

author: John Pavlus

date: September 2, 2025

source: quanta magazine https://www.quantamagazine.org/world-models-an-old-idea-in-ai-mount-a-comeback-20250902/

The latest ambition of artificial intelligence research — particularly within the labs seeking “artificial general intelligence,” or AGI — is something called a world model: a representation of the environment that an AI carries around inside itself like a computational snow globe. The AI system can use this simplified representation to evaluate predictions and decisions before applying them to its real-world tasks. The deep learning luminaries Yann LeCun (of Meta), Demis Hassabis (of Google DeepMind) and Yoshua Bengio (of Mila, the Quebec Artificial Intelligence Institute) all believe world models are essential for building AI systems that are truly smart, scientific and safe.

人工智能研究领域当前最新的雄心——特别是在那些致力于实现agi 的实验室中——是开发一种被称为 世界模型(world model) 的东西。所谓世界模型,是指 AI 在其模型内部携带的、对外部环境的表征(representation),就像一个计算机版的雪景球一样。AI 系统可以利用这种简化的内部模型,在将预测和决策应用到现实任务之前先行评估其效果。深度学习领域的几位领军人物 Yann LeCun(Meta)、Demis Hassabis(Google DeepMind)和 Yoshua Bengio(Mila,魁北克人工智能研究院)都认为,要构建真正聪明、科学、且安全的 AI 系统,世界模型是不可或缺的。

The fields of psychology, robotics and machine learning have each been using some version of the concept for decades. You likely have a world model running inside your skull right now — it’s how you know not to step in front of a moving train without needing to run the experiment first.

其实,心理学、机器人学和机器学习等领域几十年来各自一直在使用某种形式的这一概念。并且很可能,你的大脑里此刻也运行着一个世界模型——正是因为有它,你无需亲身试验就知道不该走到行驶中的火车前面。

So does this mean that AI researchers have finally found a core concept whose meaning everyone can agree upon? As a famous physicist once wrote: Surely you’re joking. A world model may sound straightforward — but as usual, no one can agree on the details. What gets represented in the model, and to what level of fidelity? Is it innate or learned, or some combination of both? And how do you detect that it’s even there at all?

那么,这是否意味着 AI 研究人员终于找到了一个所有人都认可其含义的“核心概念”(core concept)?正如一位著名物理学家曾写道:“你肯定是在开玩笑吧。” 世界模型听起来似乎很简单——但和往常一样,没有人能在细节上达成共识。世界模型中究竟该表征哪些内容?精确到什么程度?它是天生固有的,还是后天习得的,抑或两者兼有?又该如何判断这样的模型是否存在于 AI 系统之中?

It helps to know where the whole idea started. In 1943, a dozen years before the term “artificial intelligence” was coined, a 29-year-old Scottish psychologist named Kenneth Craik published an influential monograph in which he mused that “if the organism carries a ‘small-scale model’ of external reality … within its head, it is able to try out various alternatives, conclude which is the best of them … and in every way to react in a much fuller, safer, and more competent manner.” Craik’s notion of a mental model or simulation presaged the “cognitive revolution” that transformed psychology in the 1950s and still rules the cognitive sciences today. What’s more, it directly linked cognition with computation: Craik considered the “power to parallel or model external events” to be “the fundamental feature” of both “neural machinery” and “calculating machines.”

要理解这一理念,不妨追溯一下它的起源。1943 年,在“人工智能”这一术语被创造出来的前十二年,一位 29 岁的苏格兰心理学家肯尼斯·克雷克(Kenneth Craik) 出版了一本颇具影响力的专著。在书中他畅想道:“如果有机体在头脑中携带着一个关于外部现实的‘小尺度模型’……,它就能够尝试各种替代方案,选出其中最优的……,并且在各方面做出更加充分、安全而熟练的反应。”克雷克关于心理模型(mental model)或模拟的这一想法预示了 20 世纪 50 年代引发心理学范式转变的 “认知革命”(cognitive revolution) ,这场革命至今仍主导着认知科学领域。更重要的是,这一观点直接将认知与计算联系了起来:克雷克认为,“并行或模拟外部事件的能力”是无论“神经机器”还是“计算机器”都共有的“基本特征”。

肯尼斯·J·W·克雷克(Kenneth J. W. Craik) (1914–1945)是苏格兰心理学家。他于 1943 年出版的《解释的本质》(The Nature of Explanation)一书中提出了头脑中存在外部现实“微型模型”的设想,被视为心理模型概念的开端。

认知革命 (Cognitive Revolution)指 20 世纪 50 年代心理学领域从行为主义转向认知科学的一场范式转变。这场革命强调将心智视作信息处理系统,在心理学研究中引入了对大脑内部过程的关注。

The nascent field of artificial intelligence eagerly adopted the world-modeling approach. In the late 1960s, an AI system called SHRDLU wowed observers by using a rudimentary “block world” to answer commonsense questions about tabletop objects, like “Can a pyramid support a block?” But these handcrafted models couldn’t scale up to handle the complexity of more realistic settings. By the late 1980s, the AI and robotics pioneer Rodney Brooks had given up on world models completely, famously asserting that “the world is its own best model” and “explicit representations … simply get in the way.”

新兴的人工智能领域很快采用了世界模型的方法。在 1960 年代末,一个名为 SHRDLU 的 AI 系统,在一个简易的“积木世界”中回答了诸如“金字塔能支撑一个方块吗?”这样关于桌上积木的常识性问题,令观察者们惊叹不已。不过,这些手工构建的模型无法扩展以应对更逼真的环境复杂性。到了 1980 年代末,人工智能与机器人学先驱罗德尼·布鲁克斯(Rodney Brooks)干脆完全放弃了世界模型的方法,他曾著名地断言:“世界本身就是它自己的最好模型”,而“显式的内部表征……只会碍事”。

SHRDLU 是麻省理工学院计算机科学家 Terry Winograd 于 1968–1970 年间开发的一个著名人工智能程序。该程序在一个虚拟“积木世界”中执行自然语言指令,能回答关于积木世界的问题,展示了早期计算机理解人类语言并进行推理的能力。

罗德尼·布鲁克斯(Rodney Brooks) 是人工智能和机器人学领域的先驱人物之一,提出了行为式机器人学(行为层架构)理念。他在 1980 年代后期质疑使用内部世界模型的方法,提出“真实世界本身就是最好的模型”,主张让机器人通过直接感知和行动与真实环境互动,而非依赖预先构建的内部模型。

It took the rise of machine learning, especially deep learning based on artificial neural networks, to breathe life back into Craik’s brainchild. Instead of relying on brittle hand-coded rules, deep neural networks could build up internal approximations of their training environments through trial and error and then use them to accomplish narrowly specified tasks, such as driving a virtual race car. In the past few years, as the large language models behind chatbots like ChatGPT began to demonstrate emergent capabilities that they weren’t explicitly trained for — like inferring movie titles from strings of emojis, or playing the board game Othello — world models provided a convenient explanation for the mystery. To prominent AI experts such as Geoffrey Hinton, Ilya Sutskever and Chris Olah, it was obvious: Buried somewhere deep within an LLM’s thicket of virtual neurons must lie “a small-scale model of external reality,” just as Craik imagined.

直到机器学习的兴起(尤其是基于人工神经网络的深度学习),克雷克的这一“思想结晶”才重新焕发生机。深度神经网络不再依赖脆弱的手动编写规则,而是能够通过反复试错从训练环境中建立起内部的近似模型,然后利用这些模型来完成特定任务,比如驾驶虚拟赛车。最近几年,驱动 ChatGPT 等聊天机器人的大语言模型开始展现出一些未曾经过明确训练就自行涌现的能力:比如可以从一串表情符号中猜出电影的名称,甚至学会下黑白棋。对于 Geoffrey Hinton、Ilya Sutskever 和 Chris Olah 等知名的 AI 专家来说,这一切都有一个显而易见的解释:在 LLM 那密密层层的虚拟神经元深处,一定潜藏着一个“小规模的外部现实模型”,正如克雷克当初所设想的那样。

The truth, at least so far as we know, is less impressive. Instead of world models, today’s generative AIs appear to learn “bags of heuristics”: scores of disconnected rules of thumb that can approximate responses to specific scenarios, but don’t cohere into a consistent whole. (Some may actually contradict each other.) It’s a lot like the parable of the blind men and the elephant, where each man only touches one part of the animal at a time and fails to apprehend its full form. One man feels the trunk and assumes the entire elephant is snakelike; another touches a leg and guesses it’s more like a tree; a third grasps the elephant’s tail and says it’s a rope. When researchers attempt to recover evidence of a world model from within an LLM — for example, a coherent computational representation of an Othello game board — they’re looking for the whole elephant. What they find instead is a bit of snake here, a chunk of tree there, and some rope.

然而,就目前所知,事实远没有那么令人印象深刻。当今的生成式 AI 并没有学到真正的“世界模型”,看起来倒更像是学到了一个个启发式“福袋”(bags of heuristics)——大量彼此孤立的土办法,可以对特定情境做出大致回应,但无法整合成一个连贯一致的整体(其中有些规则甚至彼此矛盾)。这很像那个 “盲人摸象” 的寓言:几个盲人各自只能触摸大象的一个部分,因此都无法认识到它的全貌。一个人摸到象鼻就以为整头大象像条蛇;另一个人摸到象腿就猜测它更像一棵树;第三个人抓住大象的尾巴便断言它是一根绳子。当研究人员试图从 LLM 内部找寻世界模型的证据时——例如试图找到某种对应黑白棋棋盘的连贯计算表示——他们寻找的其实就是那头完整的“大象”。可最终,他们发现的却只是一些零散的“蛇身”碎片、一块“树干”,还有一截“绳子”。

Image

Of course, such heuristics are hardly worthless. LLMs can encode untold sackfuls of them within their trillions of parameters — and as the old saw goes, quantity has a quality all its own. That’s what makes it possible to train a language model to generate nearly perfect directions between any two points in Manhattan without learning a coherent world model of the entire street network in the process, as researchers from Harvard University and the Massachusetts Institute of Technology recently discovered.

当然,这些启发式的“碎片”并非一无是处。LLM 可以在其数万亿参数中编码无数这样的启发式“福袋”——俗话说,量变可以引发质变。正因为如此,我们甚至可以训练出一个语言模型,在未曾真正学习曼哈顿街道网络的整体世界模型的情况下,就能生成任意两点之间几乎完美的路线指引。这一点已被哈佛大学和麻省理工学院的研究人员最近的研究发现所证实。

So if bits of snake, tree and rope can do the job, why bother with the elephant? In a word, robustness: When the researchers threw their Manhattan-navigating LLM a mild curveball by randomly blocking 1% of the streets, its performance cratered. If the AI had simply encoded a street map whose details were consistent — instead of an immensely complicated, corner-by-corner patchwork of conflicting best guesses — it could have easily rerouted around the obstructions.

那么,既然几段“蛇身”“树干”和“绳索”就能完成任务,为什么还非要追求整头“大象”不可呢?用一个词来回答:鲁棒性(robustness)。当研究人员通过随机封锁曼哈顿 1% 的街道给那个擅长导航曼哈顿的 LLM 出了一道小难题时,它的性能立刻一落千丈。倘若该 AI 在内部编码的是一幅细节自洽的街道地图——而不是由无数彼此冲突的局部最优猜测七拼八凑而成的复杂拼图——它本可以轻松地绕过那些障碍。

Given the benefits that even simple world models can confer, it’s easy to understand why every large AI lab is desperate to develop them — and why academic researchers are increasingly interested in scrutinizing them, too. Robust and verifiable world models could uncover, if not the El Dorado of AGI, then at least a scientifically plausible tool for extinguishing AI hallucinations, enabling reliable reasoning, and increasing the interpretability of AI systems.

考虑到即使简单的世界模型也带来了诸多好处,我们就不难理解为什么每一家大型 AI 实验室都渴望开发它们——也不难理解为何学术界的研究人员如今也愈发热衷于世界模型。健壮且可验证的世界模型即便无法让我们找到 AGI 的“黄金之国”,也至少能提供一个在科学上行得通的工具,用于消除 AI幻觉、实现可靠的推理,并提升 AI 系统的可解释性。

埃尔多拉多 (El Dorado)原指传说中南美洲的黄金国度,引申指人们梦寐以求却难以实现的目标或终极宝藏。这里用来比喻通用人工智能(AGI)这个 AI 领域长期追寻但尚未实现的终极目标。

That’s the “what” and “why” of world models. The “how,” though, is still anyone’s guess. Google DeepMind and OpenAI are betting that with enough “multimodal” training data — like video, 3D simulations, and other input beyond mere text — a world model will spontaneously congeal within a neural network’s statistical soup. Meta’s LeCun, meanwhile, thinks that an entirely new (and non-generative) AI architecture will provide the necessary scaffolding. In the quest to build these computational snow globes, no one has a crystal ball — but the prize, for once, may just be worth the hype.

以上解释了世界模型的“what”和“why”,但至于 “how”,依然无人知晓。Google DeepMind 和 OpenAI 押注于只要有足够多的 “多模态”训练数据——比如视频、三维仿真,以及其他非文本的输入——一个世界模型就会自然而然地在 神经网络那锅统计浓汤(statistical soup) 中凝聚成形。与此同时,Meta 的 Yann LeCun 则认为,需要一种全新的非生成式 AI 架构来充当必要的支架。

在打造世界模型的征途中,谁都没有预知未来的水晶球——但至少这一次,那最终的奖赏或许真的值得所有的炒作与期待。

---

注意力还够再读一篇?

为什么2025年的ai,越来越“懂你”?揭秘chatgpt memory 运行机制

帮助“小破号”的最佳方式是点亮下面的 ♥︎ 推荐按钮