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

欢迎收听Quanta音频版。

Welcome to the Quanta audio edition.

在每两周一次的节目中，我们会为您带来一个来自Quanta的关于基础科学和数学发展的故事。

In each of these biweekly episodes, we bring you a story direct from the Quanta website about developments in basic science and mathematics.

我是苏珊·瓦莱特。

I'm Susan Vallet.

你脑子里装着一个关于世界如何运作的模型。

You're carrying around in your head a model of how the world works.

人工智能系统也需要这样做吗？

Will AI systems need to do the same?

接下来请听。

That's next.

如果你发现了一种天然的基因编辑机制，并由此引发了一场革命，会怎么样？

What if your discovery of a natural gene editing mechanism launched a revolution?

如果你的抽象数学证明与现实世界有着意想不到的联系，会怎么样？

What if your abstract math proofs had unexpected connections to the real world?

如果你的诺贝尔奖获奖研究——关于宇宙膨胀的发现——只是故事的开端呢？

What if your Nobel Prize winning work on the expanding universe was just the beginning of the story?

如果你加入我们，和史蒂夫·斯特罗加茨一起，在1月11日，通过一个播客了解所有这些以及更多内容呢？

And what if you joined us, Steve Strogatz and January 11, to learn about all of this and much more on one podcast?

收听《Quanta》杂志的《为什么的乐趣》第五季。

Listen to season five of Quanta Magazine's The Joy of Why.

从6月11日开始，每隔一个周四发布新集。

New episodes drop every other Thursday starting June 11.

人工智能研究的最新目标，尤其是在那些追求通用人工智能（AGI）的实验室中，是一种被称为“世界模型”的东西。

The latest ambition of artificial intelligence research, particularly within the labs seeking artificial general intelligence, or AGI, is something called a world model.

它是对环境的一种表征，AI将其内化，就像一个计算版的雪花玻璃球。

It's a representation of the environment that an AI carries around inside itself, like a computational snow globe.

AI系统可以利用这种简化的表征，在将其应用于现实世界任务之前，评估预测和决策。

The AI system can use this simplified representation to evaluate predictions and decisions before applying them to its real world tasks.

深度学习领域的领军人物，前Meta的杨立昆、Google DeepMind的德米斯·哈萨比斯，以及Mila（魁北克人工智能研究所）的约书亚·本吉奥，都认为世界模型对于构建真正智能、科学且安全的AI系统至关重要。

The deep learning luminaries, Jan Lakun, formerly of Meta, Demis Hassabis of Google DeepMind, and Joshua 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.

心理学、机器人学和机器学习这几个领域，几十年来都在用这个概念的不同版本。

The fields of psychology, robotics, and machine learning have each been using some version of the concept for decades.

你现在脑子里可能就有一个世界模型在运行。

You likely have a world model running inside your skull right now.

这就是为什么你知道不用先做实验，就不会走到行驶的火车前面去。

It's how you know not to step in front of a moving train without needing to run the experiment first.

那么，这是否意味着AI研究者终于找到了一个大家都能认同的核心概念呢？

So does this mean that AI researchers have finally found a core concept whose meaning everyone can agree upon?

正如一位著名物理学家曾经写过的：你肯定是在开玩笑。

As a famous physicist once wrote, surely you're joking.

世界模型听起来可能很简单，但和往常一样，没人能在细节上达成一致。

A world model may sound straightforward, but as usual, no one can agree on the details.

模型里要表示什么，精确到什么程度？

What gets represented in the model and to what level of fidelity?

它是天生的，还是学来的，还是两者兼有？

Is it innate or learned or some combination of both?

而且你怎么才能检测到它确实存在呢？

And how do you detect that it's even there at all?

了解一下这个概念的起源会有所帮助。

It helps to know where the whole idea started.

1943年，在“人工智能”这个术语被创造出来的十二年前，一位29岁的苏格兰心理学家肯尼斯·克雷克发表了一本有影响力的专著，他在书中思考道：如果生物体在头脑中携带外部现实的小型模型，它就能尝试各种选择，判断哪个最好，并以更全面、更安全、更熟练的方式做出反应。

In 1943, a dozen years before the term artificial intelligence was coined, a twenty nine year old Scottish psychologist named Kenneth Crake 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.

克雷克关于心理模型或模拟的概念，预示了20世纪50年代改变心理学的认知革命，至今仍主导着认知科学。

Craig's notion of a mental model or a simulation foreshadowed the cognitive revolution that transformed psychology in the nineteen fifties and still rules the cognitive sciences today.

更重要的是，它直接将认知与计算联系了起来。

What's more, it directly linked cognition with computation.

克雷克认为，模拟或建模外部事件的能力，是神经机制和计算机器的基本特征。

Cray considered power to parallel or model external events to be the fundamental feature of both neural machinery and calculating machines.

因此，新兴的人工智能领域热切地采用了世界建模的方法。

So the emerging field of artificial intelligence eagerly adopted the world modeling approach.

在20世纪60年代末，一个名为“希尔德鲁”的人工智能系统通过使用一个简单的积木世界来回答关于桌面上物体的常识问题，比如“金字塔能支撑积木吗？”，让观察者惊叹不已。

In the late nineteen sixties, an AI system called Shirdloo wowed observers by using a rudimentary block world to answer common sense questions about tabletop objects, like, can a pyramid support a block?

但这些手工构建的模型无法扩展到处理更复杂现实场景的程度。

But these handcrafted models couldn't scale up to handle the complexity of more realistic settings.

到了八十年代末，人工智能和机器人领域的先驱罗德尼·布鲁克斯彻底放弃了世界模型，他有一句名言：世界本身就是它最好的模型，显式的表示反而会碍事。

By the late nineteen eighties, 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.

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，才让克雷克的想法重新焕发生机。

It took the rise of machine learning, especially deep learning based on artificial neural networks to breathe life back into Craig's brainchild.

深度神经网络不再依赖脆弱的、手工编写的规则，而是通过反复试错，在内部逐步建立起对训练环境的近似理解，然后用这些理解去完成一些特定的小任务，比如驾驶虚拟赛车。

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.

过去几年，支撑ChatGPT等聊天机器人的大型语言模型开始展现出一些并非专门训练出来的涌现能力，比如根据一串表情符号猜出电影名字，或者玩奥赛罗棋。

In the past few years, 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.

结果发现，世界模型为这个谜团提供了一个合理的解释。

Turns out, World Models provided a convenient explanation for the mystery.

对于杰弗里·辛顿、伊利亚·苏茨克维和克里斯·奥拉这样的顶尖AI专家来说，这很明显。

To prominent AI experts like Jeffrey Hinton, Ilya Suttskaver, and Chris Ola, it was obvious.

在大型语言模型那密密麻麻的虚拟神经元深处，一定藏着一个对外部现实的小规模模拟，就像克雷克所设想的那样。

Buried somewhere deep within an LLM's thicket of virtual neurons must lie a small scale model of external reality, just as Quake imagined.

事实，至少就我们所知，没那么厉害。

The truth, at least so far as we know, is less impressive.

如今的生成式AI并没有世界模型，它们学到的更像是一堆经验法则，许多互不相关的规则，能对特定场景给出大致回应，但无法形成一个连贯的整体。

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 consistent whole.

实际上，有些规则甚至可能互相矛盾。

In fact, some may actually contradict each other.

这很像盲人摸象的寓言，每个盲人只摸到大象的一部分，却无法理解大象的全貌。

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.

一个人摸到象鼻子，就以为大象像条蛇。

One man feels the trunk and assumes the entire elephant is snake like.

另一个人摸到腿，就猜大象像棵树。

Another touches a leg and guesses it's more like a tree.

第三个人抓住尾巴，说大象像根绳子。

A third grasps the elephant's tail and says it's a rope.

当研究人员试图从LLM内部找到世界模型的证据，比如一个连贯的奥赛罗棋盘计算表示，他们就是在找那头完整的大象。

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.

他们找到的只是一小段蛇、一块树和一点绳子。

What they find instead is a small bit of snake here, a chunk of tree there, and some rope.

当然，这些启发式方法并非毫无价值。

Of course, such heuristics are hardly worthless.

大语言模型可以在其数万亿参数中编码无数这样的启发式。

LLMs can encode untold sackfuls of them within their trillions of parameters.

正如老话所说，数量本身就有其质量。

And as the old adage goes, quantity has a quality all its own.

这就是为什么可以训练一个语言模型，在曼哈顿任意两点之间生成近乎完美的路线，而不需要学习整个街道网络的连贯世界模型，这是哈佛大学和麻省理工学院的研究人员最近发现的。

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 MIT recently discovered.

所以，如果蛇、树和绳子的碎片就能胜任，为什么还要费心去了解大象呢？

So if bits of snake, tree, and rope can do the job, why bother with the elephant?

一句话，稳健性。

In a word, robustness.

当研究人员随机封锁了1%的街道，给这个曼哈顿导航大语言模型一个温和的曲线球时，它的表现一落千丈。

When the researchers threw their Manhattan navigating LLM a mild curveball by randomly blocking 1% of the streets, its performance cratered.

如果AI只是编码了一张细节一致的城市街道图，而不是一个由相互矛盾的猜测拼凑而成的极其复杂的街角补丁，它就能轻松绕开这些障碍。

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.

考虑到即使是简单的世界模型也能带来好处，就不难理解为什么各大AI实验室都拼命想开发它们，而学术研究人员也越来越关注对它们的研究。

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.

稳健且可验证的世界模型，即使不能揭示AGI的“黄金国”，至少也能提供一种科学上可行的工具，用于消除AI幻觉、实现可靠推理，并提高AI系统的可解释性。

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.

这就是世界模型是什么以及为什么需要它的原因。

That's the what and why of world models.

但如何构建，仍然是个未知数。

The how, though, is still anyone's guess.

谷歌DeepMind和OpenAI押注，只要有足够的多模态训练数据，比如视频、三维模拟以及其他超越纯文本的输入，世界模型就会在神经网络的统计汤中自发凝聚出来。

Google DeepMind and OpenAI are betting that with enough multimodal training data, like video, three d simulations, and other input beyond mere text, a world model will spontaneously congeal within a neural network's statistical soup.

与此同时，Meta的Lakun认为，一种全新的生成式AI架构将提供必要的支撑。

Meanwhile, Meta's Lakun thinks that an entirely new generative AI architecture will provide the necessary scaffolding.

在构建这些计算雪球的探索中，没有人能预知未来。

In the quest to build these computational snow globes, no one has a crystal ball.

但这一次，这个成果可能真的值得大家这么期待。

But the prize, for once, may just be worth the hype.

本期节目由迈克尔·卡尼翁·戈洛协助制作。

Michael Canyon Golo helped with this episode.

我是苏珊·瓦莱特。

I'm Susan Vallet.

想了解更多相关内容，请阅读约翰·帕夫拉斯在量子杂志网站quantamagazine.org上的完整文章《世界模型：AI的旧观念卷土重来》。

For more on this story, read John Pavlas' full article, World Models, An Old Idea on AI Mount a Comeback on our website, quantamagazine.org.

每周二关注本频道，收听《量子播客》。

Check out this feed every Tuesday for the Quanta Podcast.

届时，主编萨米尔·帕特尔会与我们的作者和编辑一起，聊更多《量子杂志》上最受欢迎、最有趣、最引人深思的故事。

That's where editor in chief Samir Patel talks to our writers and editors about more of Quanta's most popular, interesting, and thought provoking stories.

来自PRX。

From PRX.